SCIENCE ROBOTICS | RESEARCH ARTICLE

BIOMIMETICS Copyright © 2018

The Authors, some
Exploration of underwater life with an acoustically rights reserved;
exclusive licensee
controlled soft robotic fish American Association
for the Advancement
of Science. No claim
Robert K. Katzschmann,* Joseph DelPreto, Robert MacCurdy, Daniela Rus to original U.S.
Government Works
Closeup exploration of underwater life requires new forms of interaction, using biomimetic creatures that are
capable of agile swimming maneuvers, equipped with cameras, and supported by remote human operation. Cur-
rent robotic prototypes do not provide adequate platforms for studying marine life in their natural habitats. This
work presents the design, fabrication, control, and oceanic testing of a soft robotic fish that can swim in three di-
mensions to continuously record the aquatic life it is following or engaging. Using a miniaturized acoustic com-
munication module, a diver can direct the fish by sending commands such as speed, turning angle, and dynamic
vertical diving. This work builds on previous generations of robotic fish that were restricted to one plane in shallow
water and lacked remote control. Experimental results gathered from tests along coral reefs in the Pacific Ocean
show that the robotic fish can successfully navigate around aquatic life at depths ranging from 0 to 18 meters.
Furthermore, our robotic fish exhibits a lifelike undulating tail motion enabled by a soft robotic actuator design

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

that can potentially facilitate a more natural integration into the ocean environment. We believe that our study
advances beyond what is currently achievable using traditional thruster-based and tethered autonomous under-
water vehicles, demonstrating methods that can be used in the future for studying the interactions of aquatic life
and ocean dynamics.

INTRODUCTION top robotic fish prototypes reported in literature can swim unteth-
Problem addressed ered in three dimensions for prolonged periods of time at a range
Closeup and minimally disruptive observations of marine life are par- of depths.
ticularly useful when studying animals’ behaviors, swim patterns, and This paper describes our Soft Robotic Fish (SoFi), which builds
interactions within their habitats (1, 2). A biomimetic underwater ob- on the basic design philosophy of previous robotic fish prototypes
servatory for long-term studies could facilitate deeper understanding that we have constructed. In contrast to earlier efforts, this robot
of marine life, especially their social behaviors and how environmen- has onboard capabilities for untethered operation in ocean environ-
tal changes affect the delicate balance within the marine world. One ments, including the ability to move along three-dimensional (3D)
possibility to achieve this is using underwater vehicles that can swim trajectories by adjusting its dive planes or by controlling its buoyan-
alongside marine life to allow closeup observations. Remotely oper- cy. Onboard sensors perceive the environment, and a mission con-
ated vehicles (ROVs) or autonomous underwater vehicles (AUVs) in trol system enables a human diver to issue remote commands. SoFi
ocean environments typically use propellers or jet-based propulsion advances our previous work on soft robotic fish in several dimen-
systems (3). However, these propulsion systems generate substantial sions. The first-generation fish (10) was suspended underwater and
turbulence and have the potential to scare marine life and to prevent pneumatically actuated to swim forward at a fixed depth and to exe-
closeup observations (4). Further, the mere appearance of these vehi- cute escape maneuvers. The second-generation fish (11) used hy-
cles, typically large and rigid like a submarine, does not integrate well draulic soft actuation and incorporated dive planes for dynamic
into the marine environment. The complexity of most traditional diving. How­­­ever, the robot had limited thrust, could not withstand
ROVs also requires costly fabrication and intricate control strategies, compression at depths of more than a meter, was not able to adjust
and their large bulk restricts their tethered deployment to deeper wa- its buoy­­­ancy autonomously, and had no mechanism for under-
ter using specially equipped vessels. Smaller ROVs also generally re- water remote control and communication with a human diver. This
quire tethers, which can be cumbersome and restrict operation. paper also builds on our acoustic communication modem docu-
To address this problem, we sought to create biomimetic fish mented in (12), presenting its integration into SoFi and evaluating
robots that can be easily used by a single diver. We also wanted to its ability to enable real-­time interactive oceanic exploration. SoFi
generate lifelike undulation of the robotic fish tail for propulsion integrates and extends these previous works, achieving untethered
and to enable untethered remote control of these fish robots by a swimming and remote control at a range of depths in complex
diver. Our solution is a soft underwater robot with fluid-driven ac- environments.
tuation that swims with compliant and continuous strokes that im-
itate the movement of fish. Biomimicry potentially increases the ability Paper’s importance
of robots to approach marine life without disturbing them (5, 6) or SoFi is capable of close observations of marine life and has the poten-
their natural environment. Despite the emergence of previously un- tial to be a new platform for studying and interacting with underwater
known actuation modalities (7) that could also enable undulatory or species. It demonstrates that a soft fluidic actuator can be a successful
oscillatory biomimetic locomotion underwater (8, 9), none of the bench- propulsive mechanism for prolonged untethered underwater explo-
ration at multiple depths.
In particular, this work presents (i) a powerful hydraulic soft ac-
Distributed Robotics Laboratory, Computer Science and Artificial Intelligence Lab-
oratory, Massachusetts Institute of Technology, Cambridge, MA 02139, USA. tuator, (ii) a control mechanism that allows the robot to adjust its
*Corresponding author. Email: [email protected] buoyancy according to depth, (iii) onboard sensors to observe and

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 1 of 12

SCIENCE ROBOTICS | RESEARCH ARTICLE

record the environment, (iv) a mission control system that a human There have been various design and fabrication techniques pro-
diver can use to provide navigational commands to the robot from posed for fluidic elastomer actuators. Soft lithography (36), shape depo-
a distance using acoustic signals, and (v) extended ocean experiments sition manufacturing (37), thread-reinforced pneumatic chambers (38),
at depths ranging from 0 to 18 m. SoFi has demonstrated untethered and retractable pin casting (39) were some of the initial methods that
swimming and the ability to autonomously execute high-level com- can be used to realize soft fluid actuators. None of these methods
mands in coastal waters and coral reefs at depths of up to 18 m. In allow for the repeatable fabrication of soft fluidic actuators without
short, SoFi has the onboard capabilities of an untethered mobile un- weakening seams and integrated functional structures such as back-
derwater observatory to potentially enable nondisruptive monitoring bones. Three-dimensional printing of soft actuators and the creation
of marine life. of intelligent damping materials (40, 41) have shown that fine-grained
control of various materials allows for the automated fabrication of
Challenges in design and control heterogeneous structures with embedded liquids as functional actu-
We wanted to build and successfully deploy an untethered under- ation or passive damping channels. Although 3D printing opens pre-
water robotic fish, similar in size and behavior to living fish, that viously unknown dimensions in heterogeneous actuator design, the
can autonomously execute high-level commands received remotely materials available are not deformable and robust enough to undergo
from a diver. The challenge is to realize biomimetic swimming of a strong cyclical flexing. In the work presented here, we use monolithic
self-contained system in a compact size, with good portability, lim- casting using a lost-wax fabrication technique (11), a reliable and easily
ited power, and communication capabilities. The robotic fish has to reproducible way to fabricate soft actuators with complex inner cav-

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

execute 3D trajectories with lifelike undulatory locomotion by using ities and without seams that may compromise structural integrity.
a soft fluidic circulatory actuator and a compact buoyancy control Longevity and endurance are important challenges for self-contained
mechanism. All components of the integrated end-to-end system soft robots. Pneumatic energy sources are commonly used for the
have to be designed accordingly, including the pump, the soft actuator actuation of terrestrial soft robots (42), but external pneumatic pumps
body, the onboard control, the energy storage, the wide-view video constrain the mobility of a system, limiting autonomy and range.
camera, the onboard sensors, the acoustic communication module, Systems using a compressed air cartridge as an onboard pressure
and the remote control interface. Two major challenges related to lo- source can only operate on the order of a few minutes because of
comotion are (i) the creation of a hydraulic propulsion system that the low-energy density of compressed air and the challenge of either
can carry all crucial components needed for an untethered under- recycling or venting the air after the inflation of a cavity (10). Constant
water exploration and (ii) a low-drag design with appropriate buoy- release of gas causes nonnegligible changes in the overall buoyancy
ancy and weight distribution that can maintain structural integrity of the robotic fish, rendering depth control infeasible. In addition, a
under pressure throughout a suitable depth range. To overcome these fixed volume of gas limits deployment time. In contrast, alternately
challenges and achieve biomimetic propulsion, we had to design a transporting fluid from one chamber to another as carried out in SoFi
custom low-pressure high-flow pump and an appropriately sized soft does not require an extra storage unit, and the fluid does not need
fluidic actuator. An adjustable buoyancy unit, oil-filled chambers for to be exhausted to deflate the actuator. Using water instead of air as
electronics, custom seals, and rigid foam–filled compartments all had the transmission fluid also eases deployment underwater.
to fit within the limited volume available. Human interaction with There are multiple systems used to control the buoyancy of under­
the robot in the challenging underwater environment is also a de- water robots. The major open research problem for these mecha-
sign constraint. We created an underwater communication module nisms is reducing weight, bulk, and noise (9). One system heats and
that allows for real-time control of the robot and provides an intui- cools wax or oil to change its buoyancy (43, 44). However, this has
tive interface in a rugged, compact, and low-power package. a slow response time, especially when cooling the medium. A sec-
ond system uses a buoyancy chamber that can be filled with air or
Background and related work water; the water is pushed out of the chamber by filling it from a
Natural systems often exceed the performance of rigid robotic sys- compressed air tank (45). This system is large and requires refilling
tems because of their soft and compliant characteristics, such as the of the compressed air tanks. A third system uses electrolysis to cre-
unmatched speed and agility of a cheetah (13, 14) or the ability of a ate bubbles in a 2-ml volume (46). However, as the system is scaled
dead fish to swim upstream (15). The pioneering work in robotic fish up in size, the realizable change in volume becomes insufficient. A
was the Vorticity Control Unmanned Undersea Vehicle (16), a system fourth system, used in large underwater gliders, adjusts buoyancy by
using a driven link assembly to perform fish-like swimming. The hy- compressing or filling an air chamber and adjusts pitch by moving
draulic control of tuna fins (17) served as an inspiration to develop soft an internal mass (47, 48). Although these parts are reliable, the com-
robotic fish with hydraulic actuation, and several reviews of soft robotic plex actuation mechanisms of the plunger or bladders are intricate,
systems (18–21) have highlighted the potential advantages of deform- bulky, and difficult to scale down. A fifth mechanism, used in a batoid
able bodies for robotic systems. Several underwater vehicles using bio- robot, also compresses air through a piston. Although smaller than
inspired locomotion mechanisms have been proposed since (8). There the fourth system, it still has bulky external actuation parts (such as
have been initial steps toward soft robots that mimic fish (5, 10, 11, 22, 23), a lead screw drive) that protrude from the main body of the robot and
mantas (24–26), lamprey (27, 28), and octopi (29, 30). Several simple are difficult to incorporate in other designs such as submarines or
fish prototypes have been proposed for studying the interaction of robotic fish (49). By using similar principles to this fifth mechanism
robotic fish with real fish in small tanks (5, 6, 31–35). None of the but further miniaturizing the actuation, we designed a modular
proposed systems have demonstrated autonomous, untethered biomi- buoyancy system that is fast, simple, and effective in actuation and
metic underwater operation in a real environment at several meters of control.
depth (9). Furthermore, none of those systems have observed or inter- Underwater communication is an essential component for AUVs.
acted with aquatic life in their natural habitat. Although radio-frequency communications (50) are ubiquitous in

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 2 of 12

SCIENCE ROBOTICS | RESEARCH ARTICLE

of animals and plants. Observatories of

different levels of autonomy and biomim-
icry have been proposed. A cable car–
mounted observatory for fish assessments
(63) performed underwater stereoscopic
imaging to observe month-long small-
scale temporal patterns in fish-­habitat
interactions, but the system is difficult to
use and not suitable for most environ-
ments. A robotic fish with a Global Posi-
tioning System and a temperature sensor
demonstrated surface swimming through
Wi-Fi remote control within a small tank
(64). The aquatic hexapod AQUA (65)
is equipped with sensors to walk over ter-
rain in shallow waters. The AUV AMOUR
V (66) is a low-cost thruster-based AUV

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

capable of marine surveying and monitor-
ing. Although it carries onboard sensors,
Fig. 1. SoFi system overview. (Top, left to right) Soft robotic fish and diver interface module. (Bottom, left to right) it is less suitable for marine life observa-
Subcomponents of the system are the elastomer tail (cut view), external gear pump, BCU, two dive planes, control tions because of its disruptive thrusters.
electronics including acoustic receiver, and fisheye camera. In recent years, the development of smaller
and more maneuverable AUVs such as
biomimetic robots with sizes ranging from
a few centimeters to a meter has been a
growing field of interest (8, 9, 67). How-
ever, all of the studies are focused on dif-
ferent types of swimming locomotion and
do not demonstrate deployment in the
wild. We believe that biomimetic AUVs
have the potential for greater efficiency,
maneuverability, and stealth, which could
enable minimally disruptive environmen-
tal monitoring, proximal live fish obser-
vations, and controlled interactions with
marine life.
Fig. 2. Underwater exploration. Fish trajectory exploring a coral reef. The snapshots of the fish are equally spaced
in time by 2.6 s per representation. Contributions
This paper contributes to the field of ro-
botics with the integration of an end-to-
terrestrial applications, those signals rapidly attenuate in saltwater end system that locomotes in a biomimetic manner underwater, is
(51). Optical communications (52–54) are also challenging under- remote-controlled, and can serve as an underwater observatory for
water because they are subject to scattering and noise from ambient the study of marine life. We present a biomimetic soft robotic fish
light. We therefore used acoustic communications, which have been that is able to swim along 3D trajectories with autonomous buoy-
widely adopted for underwater applications (55–57). Although Woods ancy control to observe the biocenosis of coral reefs in the ocean.
Hole Oceanographic Institution modems (58, 59) can overcome chal- More specifically, the contributions of this work include the first
lenges such as multipath effects and Doppler shifts (60), their size and soft robotic fish prototype capable of (i) 3D controllable motion for
power consumption are too large for fish-sized robots. Similarly, other persistent operation underwater, (ii) autonomous depth control via
modems (61, 62) focus on higher data rates and longer ranges than dive planes and a miniaturized piston-­​based buoyancy control unit
required for remote-controlled operation by a diver, rendering them (BCU), (iii) underwater remote control via a miniaturized end-to-end
too bulky, expensive, and energy-consuming for our present applica- acoustic communication system, and (iv) performing at depths of
tion. Some acoustic modems (52) use hardware-defined signal gen- 0 to 18 m, as evidenced by ocean experiments.
eration and detection, but this limits available processing and reduces
versatility. Taking these works into account, we designed a lean uni-
directional communication protocol with software-defined detection RESULTS
algorithms that enable our system to send short command words We developed SoFi, a fully embedded self-contained underwater sys-
while being small and easily integrated into SoFi. tem, that swims independently and receives high-level commands
The observation of marine life using robots is particularly attractive from a human diver (Fig. 1). The robot measures 0.47 m × 0.23 m ×
when attempting to better understand the behaviors and occurrences 0.18 m, weighs 1.6 kg, is neutrally buoyant, and swims for about

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 3 of 12

SCIENCE ROBOTICS | RESEARCH ARTICLE

40 min. It propels itself by undulating its soft tail in a cyclic manner by adjusting the baseline deflection angle of the tail around which
and adjusts this undulation to swim forward or turn. The tail motion the tail undulates. The fish performed three levels of deflections in
is created by the cyclic flow of a displacement pump, and adjusting each direction, with a maximum baseline deflection of about ±30°.
the relative amount of liquid pumped into each side of the tail can Similarly, the fish could pitch its dive planes at three levels in each
generate a turning motion. Vertical swimming is achieved via dive direction, with a maximum pitch of ±45°. A sample fish trajectory
planes and a BCU. The fish is equipped with a fisheye camera at its tip along a coral reef is shown in Fig. 2, illustrating the controlled swim-
to observe its environment. An acoustic transducer is also mounted ming motion as it was commanded by a human diver. The fish changed
in front of the rigid dorsal fin, tilted upward, to receive commands direction and depth while exploring the reef, with an average swim-
from the human-operated diver interface module. ming speed of 21.7 cm/s (±3.2 cm/s) at depths of 0 to 18 m.
We performed quantitative tests in the ocean to measure the for-
Swimming along a 3D trajectory ward and turning capabilities of the fish (Fig. 3). The average swim-
The hydraulic system performed undulating tail actuation at low ming speed in a straight path was 23.5 cm/s (±0.4 cm/s), equivalent to
(0.9 Hz), medium (1.15 Hz), and high (1.4 Hz) frequencies to achieve 0.5 body lengths per second. The average turning speed was 18.3 cm/s
a range of swimming speeds. The fish executed left and right turns (±4.1 cm/s) on an average turning radius of 78.2 cm (±28.6 cm). Dy-
namic diving using the dive planes was
possible within a range of ±0.9 m from
its baseline depth at an average speed of

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

about 14.0 cm/s, equivalent to 0.8 body
heights per second. During the dive, we
changed the robot’s baseline depth with-
in 0 to 18 m by manually adjusting at-
tached weights. At deeper depths, the rigid
foam floats experienced too much com-
pression and inhibited upward diving.
Vertical diving capabilities using the
BCU were quantified in a pool. The BCU
reliably controlled depth changes of up to
2.8 m. This was repeatedly tested at dif-
ferent baseline depths of 1.6 to 2.7 m by
adjusting magnetic weights. The average
dive speed up and down was 10.6 cm/s
(±1.1 cm/s), equivalent to 0.6 body heights
per second. Commanding a step change
in depth of 0.2 m had a 10% settling time
of 17.8 s (±6.6 s). Figure 3 (top right) shows
the depth profile of a vertical dive, where
the robot was directed to continuously dive
deeper solely by compressing the air in the
piston of the BCU. The BCU responded
Fig. 3. Quantitative ocean experiments. Clockwise from top left: Straight swimming, vertical diving, left turn, and with slight oscillations around the set depth
right turn experiments of the robotic fish. until it settled and the next depth was

Table 1. Communication experiments. Cumulative results of the acoustic communication during four of the six dives, spanning 2 days and averaging about
40 active minutes per dive. Note that “steady commands” are commanded states that persisted for at least 1 s. Observations were made at an average depth of
8.1 m, a maximum depth of 18 m, a range between transmitter and receiver of 0 to 10 m, and a transmit acoustic power of 137.3 dB SPL re 1 Pa.
Dive 3 Dive 4 Dive 5 Dive 6
Total commands obeyed 67 30 111 93
Total commands missed 55 62 46 57
Steady commands obeyed 55 26 75 78
Steady commands missed 25 31 7 21
Percent of steady 68.8% 45.6% 91.5% 78.8%
commands obeyed
Fish timeouts (reversions to 63 34 81 81
neutral state)
Percent of dive spent 12.3% 8.0% 7.3% 8.1%
timed out

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 4 of 12

SCIENCE ROBOTICS | RESEARCH ARTICLE

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

Fig. 4. Underwater observatory. (A) A diver using the acoustic communication modem to remotely control the robotic fish. (B) The fish exploring complex coral reef
environments. (C) The robotic fish among marine life. (D) Pictures captured by the fish’s onboard camera.

A human diver used the diver inter-

face module to successfully steer the robot
through various complex underwater coral
reef environments. The diver commanded
levels of thrust, tail undulation frequency,
depth/pitch, and turning angle; fig. S10
illustrates the transmitted and received com-
mands for a single dive, and Table 1 sum-
marizes the detection rates. During dive 5,
the diver remained closer to the fish (with-
in a few meters), and therefore, commu-
nication was more reliable. Our analysis
focuses on “steady” commands, which were
Fig. 5. Closeup view of marine life. (Left) Onboard view filming several fish passing by the lens of SoFi. (Right) Photo not immediately followed by a different
taken by a diver from further behind, showing both the robotic fish and the observed biological fish. command within 1 s; for example, a diver
commanding a transition from lowest to
commanded. The two linear actuators within each of the two modules highest pitch would repeatedly press the “up” arrow, resulting in transient
moved symmetrically to vary the buoyancy while maintaining con- intermediate pitch states, but only the final command state is of impor-
stant pitch of the robot. Additional buoyancy control experiments are tance. If the fish did not receive any commands within a timeout period of
provided in fig. S5. about 10 s, then it would return to a neutral state and turn off the motor;
this temporary silencing would then facilitate detecting fainter commands.
Human-robot interaction In the coral reef experiments (at depths ranging from 0 to 18 m),
The human controls SoFi remotely through a custom-designed unidirec- effective communication was established at a range of up to 10 m when
tional acoustic communication modem. This system consists of the diver the robot’s motor was switched off and 5 m when the motor was switched
interface module and the acoustic receiver, both shown in Fig. 1. on. The largest factors affecting communications were environmental
We characterized the reliability of the communication modem by complexity, noise of the fish motor, and transmission distance. Ad-
transmitting a series of 200 alternating bits at a rate of 20 bits/s at ditional experiments close to the shoreline showed that, in a shallow,
varying distances and depths in a large pool. The results, shown in fig. cluttered underwater environment, the system can communicate up
S8, indicate that error-free communication was achieved at a depth of to 10 m even in the presence of motor noise.
1.8 m for ranges of up to about 15 m, with more than 97% of the data
successfully received at 21 m, and that the effective communication Oceanic exploration
range remains similar when the robot’s motor is turned on. In addition, a The robotic fish has an onboard fisheye camera that allows a remote
complete sequence of 250 16-bit data words was successfully received human operator to film the underwater exploration. This setup reduces
and decoded without errors at separations of about 0.5 m in a fish tank the impact of the diver on the marine life being filmed. The fish contin-
and 10 m in a small pool. uously operated for about 40 min during each of six ~51-min dives

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 5 of 12

SCIENCE ROBOTICS | RESEARCH ARTICLE

Fig. 7. Soft tail, pump, and BCU. (Top) Two views of the soft fish tail in an actuated

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

Fig. 6. Signal flow within the system. The command flow from a human diver to state (left) and custom external gear pump in an exploded view (right). (Bottom) One
the robot. The diver sends acoustic commands such as thrust, left/right, and up/ of two identical BCU modules in an exploded view.
down as well as camera modes from the transmitter within the diver interface
module. The analog signal travels several meters underwater and is then amplified
by the receiver and parsed by the microcontroller. The microcontroller adjusts the
eration. Whereas most soft robots are pneumatically powered and
pump speed, the dive plane position, the BCU, and the mode of the camera.
tethered, our hydraulically driven soft actuator enables prolonged un-
tethered swimming over several hundred meters for 40 min. The hy-
over 3 days, accumulating 240 min of controlled exploration (table S2). draulic system can perform low- to high-frequency tail actuation to
Each dive used a single charge of a 35–watt-hour battery. The average achieve a range of swimming speeds and can execute turns by adjust-
depth was 8.1 m, and the maximum depth was 18 m. We recorded ing the baseline deflection around which the tail undulates. Dive planes
the number of tail strokes performed for an average dive (table S3). and the BCU enable vertical swimming. A camera is mounted at the tip
On the basis of the average swimming speed observed duri­ng the of the robot, allowing a diver to remotely explore and capture closeup
quantitative tests, for a single dive, this would correspond to 296.8 m recordings of marine life and environments.
(±5.1 m) of straight swimming, 17.6 m (±3.9 m) of turning left, and The acoustic communication system provides a compact, software-­
14.6 m (±3.3 m) of turning right. The ocean experiments showed that defined modulation scheme for transmitting data that is robust to
the robotic fish is capable of 3D controllable motions in natural envi- substantial noise and interference from complex environments. In
ronments, in the presence of currents. open-ocean coral reefs including obstacles, sources of noise, and
During the ocean experiments, we also made preliminary observa- multipath effects, SoFi was able to transmit 16-bit words once per
tions of the robot’s ability to make closeup observations of marine second over distances of up to 10 m. This successfully enabled divers
life and record their responses. Without the intent to imitate any to send high-level commands and navigate the robotic fish, observ-
specific species of biological fish, the robotic fish was colored white ing the marine life and exploring their surroundings.
except for the black lens and acoustic transducer (table S1). There
were several closeup encounters with underwater life during which Limitations and future steps
SoFi observed fish and their interactions. We made the personal ob- We demonstrated that SoFi can navigate in natural environments. The
servation that SoFi did not cause other fish to flee upon these close next steps are to use SoFi as an instrument to (i) study the behavior of
encounters, even when within less than a meter. Figure 4 shows ex- marine life over long periods of time without human interference with
amples of high-level remote control by a human diver and depicts the scene, (ii) study whether SoFi can be used to influence the behavior
some of SoFi’s explorations and observations: exploring complex coral of marine life, and (iii) create robotic swarms. These research direc-
reef environments, encountering schools of fish, and capturing im- tions are enabled by SoFi and are the subject of future work.
ages with the onboard camera. Figure 5 and movie S1 also show SoFi SoFi can be created at different scales, but its swimming behavior
approaching overhanging reefs and other environments with fish swim- depends on its size. Smaller robotic fish can barely overcome ocean
ming nearby. As the movie shows, multiple fish swim parallel to the currents and need external power (68), whereas larger robotic fish are
robot a few centimeters below it and also pass a few centimeters in front more difficult to prototype and to handle by a diver. SoFi can currently
of its lens. The fish did not appear to change their swimming trajectory swim up to 0.51 body lengths per second, which is comparable to
as SoFi approached them in these cases, suggesting that SoFi has the other robotic fish prototypes (69, 70) but still leaves room to improve
potential to integrate into the natural underwater environment. toward real fish capabilities of 2 to 10 body lengths per second (71, 72).
Further optimizations of the pump system, the tail geometries, and
the exterior profile of SoFi may improve swimming efficiency.
DISCUSSION The dive planes provide only fine-tuned control at a limited depth
Conclusion range. Once the range is exceeded, the compression of the fish’s flota-
We have presented an untethered soft-bodied robotic fish that demon- tion becomes so strong that inverting the pitch of the dive planes will
strates prolonged and consistent underwater remote-controlled op- not allow returning to the original depth; the diver must manually

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 6 of 12

SCIENCE ROBOTICS | RESEARCH ARTICLE

pects should be taken into account when

designing robot-fish studies with our pro-
totype in the future.
In contrast to previous robot-fish stud-
ies, the robot prototype presented in this
work provides the opportunity to perform
studies of the biocenosis of coral reefs and
other marine environments within nat-
ural habitats. In the future, researchers
could use the soft robotic fish described
in this paper and easily change its size,
color, and shape to emulate various types
of fish with different dynamic behaviors.
The integrated camera and the ability to
remotely control the robot in three dimen-
Fig. 8. Acoustic communication. (Left) Exploded view of the diver interface module, containing the transmitter.
sions at a variety of depths allow the sys-
(Right) Schematic view of the transmitter and receiver pipelines. tem to observe and approach marine life.

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

The soft fish presented in this paper
can also be rapidly fabricated to create a
swarm of robotic fish. Such a swarm could
adjust the weight during the dive to change to another depth range. enable studies of schools of fish and their interactions in the presence
Using asynchronous control of the BCU modules would enable in- of varying ocean dynamics (82–85).
creased pitch control (see fig. S4 for details), although the BCU is still
limited in its diving speed and range. The speed and range can be in-
creased by enlarging the body to allow for larger dive planes or BCU MATERIALS AND METHODS
pistons while balancing against the trade-off of increased drag. Up- The objective of this study is to show that we can create a soft robot-
grading the tail design to four instead of two fluidic chambers, with ic fish that uses undulating motion to swim in the ocean and explore
one chamber per quadrant, would also allow steering in the vertical underwater life and structures.
direction through biased undulation of the tail in the vertical plane.
Improving the acoustic modem could allow a diver to be further System architecture
away. Optimizing the modulation parameters, implementing dif- The full system and its major subcomponents are shown in Fig. 1.
ferent protocols such as frequency hopping, refining the transducer The nose of the fish is a waterproof housing for the fisheye camera,
and amplifier circuitry, and reducing the motor noise could increase microcontroller, computer, motor driver, wireless communica-
data rates and detector robustness. In addition, the modem can be tion electronics, inertial measurement unit, and depth sensor. The
extended to control multiple robots or to be bidirectional and pro- housing is 3D-printed and waterproofed by brush-coating it with
vide the diver with real-time feedback. preheated epoxy paint and subsequent degassing. Behind the nose
The integrated camera enables more autonomous surveying capa- is the dive plane assembly, consisting of two individually control-
bilities (73). Monocular self-localization would enable the fish to build lable dive plane units. Each unit consists of a dive plane directly
maps of the underwater environment and explore it further. Instead mounted onto the lever arm of a waterproof servo motor. The dive
of using acoustic communication for lower-level settings such as plane assembly is mounted to the end of the gear pump’s DC mo-
thrust and depth, a diver could remotely command higher-level mis- tor. The motor and gear pump unit are directly attached to the soft
sion parameters such as regions to explore or specific marine life to fish tail. Underneath the gear pump motor is a lithium polymer bat-
follow. tery to power all components. Above the gear pump is the BCU.
There are many potential future applications of the robotic fish The mass of the complete assembly was slightly adjusted to make
described in this paper in the emerging field of ethorobotics (1, 2). We it almost neutrally buoyant using internal rigid urethane foam
are inspired by previous work that considered robot-animal inter- chambers and additional magnetic weights placed underneath the
actions, including research on robot-cockroach societies (74), remote- robotic fish.
controlled cow gathering (75), pet care robots (76), honeybee robots The flow of commands within the system is depicted in Fig. 6. The
(77), and guinea fowl (78). More recently, studies in small fish tanks diver commands a change to the fish state via the gamepad controller
began to specifically investigate interactions between robotic fish lures within the diver interface module. The command is encoded into an
and natural fish, such as golden shiners (5, 6), zebrafish (31–33, 79, 80), acoustic signal transmitted via the acoustic transducer to the amplify-
or Siamese fighting fish (35). This previous work considers controlled ing receiver within the fish. The microcontroller decodes the re-
studies in laboratory environments, conducted in tanks with unac- ceived command and adjusts its state accordingly. Changes to the
tuated fish replicas or primitive robotic fish prototypes with servo-­ swimming speed or turning motion change the behavior of the dis-
actuated tails. These previous studies showed that the appearance or placement pump and therefore the soft tail undulation. Changes in
biomimetic locomotion of the robotic device does not ensure integra- pitch or depth, depending on the mode, are sent to the servos of the
tion within a school of fish because acceptance depends on multiple dive planes or the BCU. Changes to the video recording state are for-
signals. It was also found that a robotic fish can be differently per- warded to the single-board computer, which records from the fisheye
ceived in terms of attractiveness by real fish (32, 35, 79, 81). These as- camera.

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 7 of 12

SCIENCE ROBOTICS | RESEARCH ARTICLE

Soft body for undulating locomotion ducing sufficiently large cyclic body deflections and the least acoustic
The fish achieves undulating locomotion via a hydraulically actuated noise. An external gear pump design produced the largest body de-
soft fish tail with two internal cavities. The soft fish tail, shown in flection and therefore the best swimming performance but consumed
Fig. 7, is a fluidic elastomer actuator (39, 86, 87). The design mimics an order of magnitude more power and produced higher noise levels.
the rear portion of a fish, encompassing the posterior peduncle and A detailed study of the various actuation systems is provided in (90). We
the caudal fin. The tail can continuously bend along its vertical center chose an external gear pump (Fig. 7) for the fully integrated robotic
constraint layer by fluidic actuation of two lateral cavity structures. fish because of its better swimming performance, lower part count,
The inextensible and stiffer center constraint layer splits the tail evenly and easier controllability.
along a vertical plane. An actuator consists of evenly spaced ribs The motor controller operates the motor attached to the pump
with hollow sections in between, connected by a center channel and through a trapezoidal voltage profile, alternating from positive to
accessible by a front inlet. The rib structure allows for expansion or negative voltages after each half-cycle. This profile rotates the motor
contraction of the thin exterior skin under positive or negative fluidic shaft back and forth, causing the pump to create a cyclic hydraulic
pressure, respectively. These expanding or contracting motions bend flow. Asymmetrically varying the flow intensity for each half-phase
the inextensible center constraint layer. The rib structure is evenly can enable yaw control by creating a pressure bias in the tail.
spread along the fin, leading to a continuous flexing of the whole
body under fluidic pressure. The inherent elasticity of the body forces Depth control
it back into its neutral state after each pulse of actuation. A fluidic The depth of SoFi is controlled by dive planes or the BCU. The dive

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

flow alternating into each lateral cavity structure leads to a complex planes, shown in Fig. 1, allow the diver to finely control the robot’s
undulating motion of the soft body and enables swimming. change in depth through dynamic diving for limited deviations from
The fabrication of the soft body with its integrated constraint its baseline depth before a buoyancy adjustment is needed. Manually
layers and posterior fins is realized through a lost-wax fabrication adding or removing magnetic weights attached to the bottom of the
process. The interior cavity of the body is realized through a lost-wax robot adjusts the neutral depth level. This allows the diver to operate
core (fig. S1). The constraint layers are laser-cut, and the surrounding the robot over a larger depth range.
molds are 3D-printed. The wax cores and the constraint layers, which In addition, the diver can remotely adjust the neutral buoyancy
also act as the posterior fins, are assembled together into a mold (fig. of the robot using the BCU. The BCU, shown in Fig. 1, can simulta-
S2). The soft silicone elastomer and low-density glass bubbles are neously control the buoyancy and pitch of the robot. The mechanical
mixed at a mass ratio of 40:1 to achieve a desired rubber density of design of the BCU comprises two mirroring volume control modules
0.94 g/cm3. This mix is filled into the cavity and allowed to cure. Heat- in the form of two pistons. The BCU is symmetrically oriented at
ing the resulting body in an oven and then in a water bath removes the the robot’s center of buoyancy. An exploded view of a single unit is
interior wax body and creates the soft body. presented in Fig. 7. A single unit contains a micro linear actuator
with potentiometer feedback (PQ12, Actuonix, Victoria, Canada) that
Cyclic hydraulic actuation sits within a watertight cylinder and moves a piston. A closed-loop
The soft tail is actuated by a hydraulic pump at a desired undulation proportional-integral-derivative (PID) controller with pressure feed-
frequency and amplitude. The outlets of the hydraulic pump are di- back from an integrated pressure sensor is used to drive the volume-­
rectly attached to the soft body to allow for water movement between changing actuators. Ascent, descent, and hovering can be achieved
the two inner cavities in a closed-loop fashion. Alternating the flow over several meters by symmetrically controlling the pistons. The pitch
direction leads to a flexing actuation of the soft body in a side-to-side can also be modified by asymmetrically controlling the two pistons
manner, propelling the robot forward. The soft tail has removable (fig. S4).
plugs at the caudal fin, which are initially removed, so water can fill The BCU’s performance was quantitatively evaluated in an indoor
the actuation chambers by running the self-priming gear pump at a swimming pool with a depth of 4.2 m. There were no substantial dis-
low frequency for a short duration. After all air has been removed, the turbances in the environment except for the pool circulation and
plugs are inserted to seal the chambers. swimmers in adjacent lanes. The gains of the PID controller were es-
We dimensioned the custom-designed pump (fig. S3) and its at- timated by averaging the results of two frequency response tests (91).
tached motor based on the maximum pressure required and the vol- Before the start of each trial, the robot’s weight was adjusted for neu-
umetric flow rate. We estimated the effective volumetric flow rate tral buoyancy at a desired baseline depth. Desired depth values were
based on the displaced volume of fluid for a single static deflection and then commanded as a step function. Once a set depth was held for 4 s
the desired flapping frequency. Initial values for the desired flapping within an error margin of 10%, the next depth level was commanded.
frequency and the amplitude of the soft tail were determined based The robot’s microcontroller continuously logged depth by reading a
on previous studies on self-propelling foils driven by an external ac- pressure sensor. Each run started at a different depth to investigate vary-
tuator (88, 89). A custom pump, its attached motor, and a waterproof ing baselines. We measured the depth, speed, duty cycles, and error.
housing were then specified, designed, and built. The effectiveness
of six different self-contained designs based on a centrifugal pump, a Underwater communication
flexible impeller pump, an external gear pump, and rotating valves was Acoustic modem design
compared. These hydraulic actuation systems combined with the soft We designed a compact unidirectional acoustic communication mo-
tail were then measured at low and high oscillation frequencies. The dem to allow SoFi to support remote-controlled operation. The diver
propulsive force, deflection characteristics of the soft tail, acoustic interface module (Fig. 8, left) contains the transmitter and allows a
noise of the pump, and overall efficiency of the system were recorded. diver to issue commands, whereas the receiver is embedded within
A brushless, centrifugal pump combined with a custom-printed ro- SoFi’s head. Tight volumetric constraints made accommodating
tating valve performed most efficiently at both test frequencies, pro- existing underwater modem designs impractical. Thus, we implemented

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 8 of 12

SCIENCE ROBOTICS | RESEARCH ARTICLE

a new low-power, low-cost, software-defined acoustic modem, rep- using a [15,11] Hamming encoding with an additional parity bit.
resented schematically in Fig. 8 on the right and described in detail This vocabulary of commands can then be used to remotely control
in (12). the fish.
The acoustic modem’s transmitter is housed in the diver interface Acoustic modem testing
module, which incorporates an oil-filled rigid outer shell (22 cm by We evaluated the acoustic modem in a pool, a fish tank, and the ocean.
22 cm by 6 cm) with a transparent flexible membrane on one face. The The system was first evaluated in a tank (1.2 m × 0.3 m × 0.45 m)
membrane, a soft cast-molded silicone rubber (SORTA-Clear 40, Smooth- and pool (23 m × 12.5 m × 2.2 to 4.2 m) to test the modem under
On), retains nonconductive mineral oil within the housing and al- controlled conditions. These environments facilitate multipath re-
lows for pressure equalization underwater. The flexibility and molded flections due to the enclosed configuration, hard walls, and shallow
shape of the membrane allow the control buttons within the module to depth, approximating the types of interference observed in open-ocean
be pressed by the diver when selecting a desired fish state. These com- deployments.
mands are read by a Raspberry Pi single-board computer via Uni- As described in more detail in the Supplementary Materials, tests
versal Serial Bus and are encoded as a specific sequence of ultrasonic were performed during development to choose parameters of the
acoustic tones, which are then converted to audio signals by a digital- modulation scheme and decoding algorithm. Then, to evaluate the
to-analog converter (HiFiBerry). The analog signals are amplified communication reliability of the completed modem, we transmitted
via a Class G differential audio amplifier (MAX9788) and are then a series of 200 alternating bits at a rate of 20 bits/s over a sequence of
impedance-­matched to the output ceramic transducer (Aquarian Sci- increasing distances and depths. For each transmission, the percent-

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

entific AS-1 hydrophone) via a step-up transformer (Pico Electronics age of bits correctly decoded by the receiver and the longest error-free
32146). The hydrophone has a transmit sensitivity of 116 dB relative segment of received bits were extracted. After evaluating single-bit
to (re) 1 V/Pa (1 voltage root mean square input at 1-m range) at transmissions, transmissions of complete words were investigated by
30 kHz and was driven at 32.8 V peak-to-peak, yielding a transmit sending a predefined series of 250 16-bit data words using 50 ms for
sound pressure level (SPL) of 137.3 dB re 1 Pa. each bit and 200 ms between words. The correctness of the decoded
The modem’s receiver, housed within an oil-filled chamber in sequence was then measured. Last, the complete modem integrated
SoFi’s head (see Fig. 1), occupies less than 30 cm3. Audio signals are within SoFi was used in the open ocean to evaluate performance in
transduced by a hydrophone (Aquarian Scientific AS-1) with a voltage-­ real-world operations.
mode receive sensitivity of −207 dB re 1 V/Pa, amplified and filtered
by a custom Junction Field Effect Transistor preamplifier with a 17-dB Open-ocean experiments
gain, filtered and amplified by a band-pass filter with a 20- to 40-kHz We tested the complete system in the open ocean, with a diver re-
passband and a 40-dB gain, and digitized by an Mbed microcontrol- motely adjusting the fish’s state and navigating it to points of interest
ler. A variable-gain amplifier controlled by the Mbed allows dynamic in a complex underwater environment. Six dives were conducted
signal equalization with a gain from 0 to 40 dB. Modulation and over the course of 3 days, exploring the Somosomo Strait in Taveuni,
demodulation are both defined in software for versatility, facilitating Fiji (see table S2 for details). This location offers numerous coral reef
alternate modulation protocol implementations. The receiver con- environments with varying tidal conditions, allowing SoFi to be eval-
sumes 815 mW, with the Mbed using about 740 mW of that power. uated in real-world conditions where the interactions of marine life
Communication frequencies were chosen by considering typical and the biocenosis of coral reefs can be studied.
ranges of human hearing, frequency-dependent attenuation in un- The robot conducted about 40 min of continuous observation
derwater channels (56), Doppler effects, SoFi’s motor noise, the mi- during each dive, totaling about 240 min of controlled exploration
crocontroller’s sampling capabilities, parameters of the receiver’s at an average depth of 8.1 m and a maximum depth of 18 m. We per-
detection algorithm, expected sources of environmental noise such as formed an additional 90 min of preparatory swim tests in shallow
wind and waves (92, 93), and marine life. Noise produced by fish is ocean waters to test the control system, communication, and video
typically below 10 kHz (94), and the hearing ranges of common recording. All of these tests evaluated the effectiveness of SoFi’s bio-
aquatic species decay significantly above 10 kHz (95, 96), although mimetic actuation and the usability of the acoustic communication
some cetaceans and pinnipeds can hear well above this range (4). interface for remote control. The distance between the operator and
Taking into account all of these considerations, 36 kHz was chosen SoFi was typically between 1 and 10 m, and the transmit power of
for a logical 0 and 30 kHz was chosen for a logical 1. the acoustic modem was 137.3 dB SPL re 1 Pa. The robot’s trajec-
Considering the design constraints, a modulation scheme that tories along the reefs and following other fish were documented by
could be efficiently implemented in software on a microcontroller two or more divers using GoPro Hero 3, Canon PowerShot S100, and
while still being robust to multipath effects and Doppler shifts was Olympus Tough TG-1 cameras from a distance of several meters.
designed. It uses pulse-based frequency-shift keying and a compu- Qualitative observations were made during five of the dives, during
tationally efficient software-defined demodulation approach leverag- which SoFi explored the coral reef environments. The magnetic
ing the Goertzel algorithm (97) and a custom dynamic peak detection weights were adjusted at the beginning of each dive for neutral buoy-
algorithm. The chosen parameters support 2048 distinct messages ancy and then the robot was continuously operated via the acoustic
with a data rate of one message per second at 20 bits/s. Further details modem. The distance between controller and robot was varied to
on the algorithm can be found in (12). understand the effective communication range. The fish was steered
The desired fish state, encoded as a 16-bit word, is transmitted throughout the coral reefs, going as close as possible to interesting
from the controller once per second. Each command describes a environmental features and marine life. Such dives provided qualita-
desired state of the fish including tail oscillation frequency (2 bits), tive observations of SoFi’s swimming capabilities in constrained and
oscillation amplitude (2 bits), pitch or depth (3 bits), yaw (3 bits), and unconstrained areas, of the acoustic communication reliability, and
video recording (1 bit). These 11 bits are expanded to a 16-bit word of the effect that SoFi has on nearby fish.

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 9 of 12

SCIENCE ROBOTICS | RESEARCH ARTICLE

In addition, one dive was dedicated to performing quantitative 5. S. Marras, M. Porfiri, Fish and robots swimming together: Attraction towards the robot
demands biomimetic locomotion. J. R. Soc. Interface 9, 1856–1868 (2012).
swimming tests on the ocean floor at a baseline depth of about 7 m.
6. S. Butail, N. Abaid, S. Macrì, M. Porfiri, Fish–robot interactions: Robot fish in animal
We installed several premeasured ropes to define a reference volume behavioral studies, in Robot Fish, R. Du, Z. Li, K. Youcef-Toumi, P. Valdivia y Alvarado, Eds.
(4 m × 4 m × 1 m) for measuring and filming the robot’s ability to (Springer, 2015), Springer Tracts in Mechanical Engineering, pp. 359–377.
swim straight, turn right, turn left, dive up, and dive down. We per- 7. L. Hines, K. Petersen, G. Z. Lum, M. Sitti, Soft actuators for small-scale robotics. Adv. Mater.
formed three trials for each ability. During all trials, the thrust was 29 (2017).
8. W.-S. Chu, K.-T. Lee, S.-H. Song, M.-W. Han, J.-Y. Lee, H.-S. Kim, M.-S. Kim, Y.-J. Park,
set to maximum and the undulation frequency was set to medium K.-J. Cho, S.-H. Ahn, Review of biomimetic underwater robots using smart actuators.
(1.15 Hz). For right and left turns, the yaw was set to ±30° and the dive Int. J. Precis. Eng. Manuf. 13, 1281–1292 (2012).
planes were set to neutral. For up or down swimming, the dive planes 9. A. Raj, A. Thakur, Fish-inspired robots: Design, sensing, actuation, and autonomy—A
were adjusted to ±45° and yaw was set to neutral. Yaw and pitch were review of research. Bioinspir. Biomim. 11, 031001 (2016).
10. A. D. Marchese, C. D. Onal, D. Rus, Autonomous soft robotic fish capable of escape
both neutral for straight swimming. At the beginning of each trial, a
maneuvers using fluidic elastomer actuators. Soft Robot. 1, 75–87 (2014).
diver repositioned the fish to its starting position at the center of one 11. R. K. Katzschmann, A. D. Marchese, D. Rus, Hydraulic autonomous soft robotic fish for 3D
of the bounding planes of the reference volume and then released the swimming, in Experimental Robotics: The 14th International Symposium on Experimental
fish without pushing it. This diver also took notes during trials. A Robotics, M. A. Hsieh, O. Khatib, V. Kumar, Eds. (Springer International Publishing, 2014),
second diver commanded the desired fish state from the starting po- vol. 109, pp. 405–420.
12. J. DelPreto, R. K. Katzschmann, R. MacCurdy, D. Rus, A compact acoustic communication
sition. Two additional divers filmed the trials from the side and top, module for remote control underwater, in WUWNET ’15: Proceedings of the 10th
standing or floating at the boundary of the reference volume. International Conference on Underwater Networks & Systems (ACM, 2015), pp. 1–7.

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

Throughout all sessions, the diver interface module transmitted 13. A. M. Wilson, J. C. Lowe, K. Roskilly, P. E. Hudson, K. A. Golabek, J. W. McNutt, Locomotion
the desired fish state once per second using a bitrate of 20 bits/s. By dynamics of hunting in wild cheetahs. Nature 498, 185–189 (2013).
14. S. Seok, A. Wang, M. Y. Chuah, D. J. Hyun, J. Lee, D. M. Otten, J. H. Lang, S. Kim, Design
recording logs of commands on both the transmitter and receiver, the
principles for energy-efficient legged locomotion and implementation on the MIT
percentage of commands successfully received and executed by SoFi cheetah robot. IEEE ASME Trans. Mechatron. 20, 1117–1129 (2015).
could be extracted. In addition, qualitative observations regarding 15. D. N. Beal, F. S. Hover, M. S. Triantafyllou, J. C. Liao, G. V. Lauder, Passive propulsion in
achievable communication distances, the effect of real-world obsta- vortex wakes. J. Fluid Mech. 549, 385–402 (2006).
cles such as coral reefs on transmission reliability, the effect of ambi- 16. J. M. Anderson, N. K. Chhabra, Maneuvering and stability performance of a robotic tuna.
Integr. Comp. Biol. 42, 118–126 (2002).
ent noise such as from marine life, and the effect of the system on 17. V. Pavlov, B. Rosental, N. F. Hansen, J. M. Beers, G. Parish, I. Rowbotham, B. A. Block,
surrounding organisms were made. Hydraulic control of tuna fins: A role for the lymphatic system in vertebrate locomotion.
We estimated the number of tail strokes at various combinations Science 357, 310–314 (2017).
of thrust and yaw using the logs of the executed fish states. Using the 18. D. Rus, M. T. Tolley, Design, fabrication and control of soft robots. Nature 521, 467–475
(2015).
highest commanded undulation frequency, we estimated a conserva-
19. D. Trivedi, C. D. Rahn, W. M. Kier, I. D. Walker, Soft robotics: Biological inspiration, state of
tive total stroke duration. We then weighted according to thrust level the art, and future research. Appl. Bionics. Biomech. 5, 99–117 (2008).
and used the average speed from the quantitative tests to estimate the 20. S. Kim, C. Laschi, B. Trimmer, Soft robotics: A bioinspired evolution in robotics. Trends
swimming distance. Biotechnol. 31, 287–294 (2013).
21. C. Laschi, B. Mazzolai, M. Cianchetti, Soft robotics: Technologies and systems pushing the
boundaries of robot abilities. Sci. Robot. 1, eaah3690 (2016).
22. A. D. Marchese, C. D. Onal, D. Rus, Towards a self-contained soft robotic fish: On-board
SUPPLEMENTARY MATERIALS
pressure generation and embedded electro-permanent magnet valves, in Experimental
robotics.sciencemag.org/cgi/content/full/3/16/eaar3449/DC1
Robotics: The 13th International Symposium on Experimental Robotics, J. P. Desai, G. Dudek,
Materials and Methods
O. Khatib, V. Kumar, Eds. (Springer International Publishing, 2013), vol. 88 of Springer
Fig. S1. Wax core fabrication.
Tracts in Advanced Robotics, pp. 41–54.
Fig. S2. Tail fabrication.
23. P. Phamduy, M. Vazquez, A. Rizzo, M. Porfiri, Miniature underwater robotic fish for
Fig. S3. External gear pump.
animal-robot interactions, in ASME 2016 Dynamic Systems and Control Conference (ASME,
Fig. S4. Buoyancy control system.
2016), vol. 2, p. V002T17A009.
Fig. S5. Additional buoyancy control experiments.
24. K. Suzumori, S. Endo, T. Kanda, N. Kato, H. Suzuki, A bending pneumatic rubber actuator
Fig. S6. Acoustic reflections.
realizing soft-bodied manta swimming robot, in International Conference on Robotics and
Fig. S7. Motor’s broad spectrum noise.
Automation (IEEE, 2007), pp. 4975–4980.
Fig. S8. Acoustic range tests.
25. A. Cloitre, B. Arensen, N. M. Patrikalakis, K. Youcef-Toumi, P. V. Y. Alvarado, Propulsive
Fig. S9. Performance of tone detection algorithm.
performance of an underwater soft biomimetic batoid robot, in The Twenty-fourth
Fig. S10. Ocean communication tests.
International Ocean and Polar Engineering Conference (International Society of Offshore
Table S1. Color measurements of exposed parts.
and Polar Engineers, 2014), vol. 3, pp. 326–333.
Table S2. Dive summaries.
26. T. Li, G. Li, Y. Liang, T. Cheng, J. Dai, X. Yang, B. Liu, Z. Zeng, Z. Huang, Y. Luo, T. Xie,
Table S3. Tail strokes.
W. Yang, Fast-moving soft electronic fish. Sci. Adv. 3, e1602045 (2017).
Movie S1. Underwater experiments.
27. C. Stefanini, S. Orofino, L. Manfredi, S. Mintchev, S. Marrazza, T. Assaf, L. Capantini,
E. Sinibaldi, S. Grillner, P. Wallén, P. Dario, A novel autonomous, bioinspired swimming
robot developed by neuroscientists and bioengineers. Bioinspir. Biomim. 7, 025001
REFERENCES AND NOTES (2012).
1. J. Krause, A. F. T. Winfield, J.-L. Deneubourg, Interactive robots in experimental biology. 28. L. Manfredi, T. Assaf, S. Mintchev, S. Marrazza, L. Capantini, S. Orofino, L. Ascari,
Trends Ecol. Evol. 26, 369–375 (2011). S. Grillner, P. Wallén, Ö. Ekeberg, C. Stefanini, P. Dario, A bioinspired autonomous
2. Á. Miklósi, L. Gerencsér, Potential application of autonomous and semiautonomous swimming robot as a tool for studying goal-directed locomotion. Biol. Cybern. 107,
robots in the study of animal behaviour, in 2012 IEEE 3rd International Conference on 513–527 (2013).
Cognitive Infocommunications (CogInfoCom) (IEEE, 2012), pp. 759–762. 29. M. Calisti, M. Giorelli, G. Levy, B. Mazzolai, B. Hochner, C. Laschi, P. Dario, An octopus-
3. R. Bogue, Underwater robots: A review of technologies and applications. Ind. Rob. 42, bioinspired solution to movement and manipulation for soft robots. Bioinspir. Biomim. 6,
186–191 (2015). 036002 (2011).
4. B. L. Southall, A. E. Bowles, W. T. Ellison, J. J. Finneran, R. L. Gentry, C. R. Greene Jr., 30. F. G. Serchi, A. Arienti, I. Baldoli, C. Laschi, An elastic pulsed-jet thruster for Soft Unmanned
D. Kastak, D. R. Ketten, J. H. Miller, P. E. Nachtigall, W. J. Richardson, J. A. Thomas, Underwater Vehicles, in 2013 IEEE International Conference on Robotics and Automation
P. L. Tyack, Overview. Aquat. Mamm. 33, 411–414 (2007). (ICRA) (IEEE, 2013), pp. 5103–5110.

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 10 of 12

SCIENCE ROBOTICS | RESEARCH ARTICLE

31. G. Polverino, N. Abaid, V. Kopman, S. Macrì, M. Porfiri, Zebrafish response to robotic fish: 59. L. Freitag, M. Johnson, M. Grund, S. Singh, J. Preisig, Integrated acoustic communication
Preference experiments on isolated individuals and small shoals. Bioinspir. Biomim. 7, and navigation for multiple UUVs, in OCEANS 2001 MTS/IEEE Conference and Exhibition
036019 (2012). (IEEE, 2001), vol. 4, pp. 2065–2070.
32. V. Cianca, T. Bartolini, M. Porfiri, S. Macrì, A robotics-based behavioral paradigm to 60. I. F. Akyildiz, D. Pompili, T. Melodia, Underwater acoustic sensor networks: Research
measure anxiety-related responses in zebrafish. PLOS ONE 8, e69661 (2013). challenges. Ad Hoc Netw. 3, 257–279 (2005).
33. T. Ruberto, V. Mwaffo, S. Singh, D. Neri, M. Porfiri, Zebrafish response to a robotic replica 61. M. S. Martins, N. Pinto, G. Rocha, J. Cabral, S. L. Mendez, Development of a 1 Mbps low
in three dimensions. R. Soc. Open Sci. 3, 160505 (2016). power acoustic modem for underwater communications, in 2014 IEEE International
34. F. Bonnet, Y. Kato, J. Halloy, F. Mondada, Infiltrating the zebrafish swarm: Design, Ultrasonics Symposium (IUS) (IEEE, 2014), pp. 2482–2485.
implementation and experimental tests of a miniature robotic fish lure for fish–robot 62. A. Sánchez, S. Blanc, P. Yuste, A. Perles, J. J. Serrano, An ultra-low power and flexible
interaction studies. Artif. Life Robot. 21, 239–246 (2016). acoustic modem design to develop energy-efficient underwater sensor networks.
35. D. Romano, G. Benelli, E. Donati, D. Remorini, A. Canale, C. Stefanini, Multiple cues Sensors 12, 6837–6856 (2012).
produced by a robotic fish modulate aggressive behaviour in Siamese fighting fishes. 63. P. Fischer, A. Weber, G. Heine, H. Weber, Habitat structure and fish: Assessing the role of
Sci. Rep. 7, 4667 (2017). habitat complexity for fish using a small, semiportable, 3-D underwater observatory.
36. Y. Xia, G. M. Whitesides, Soft lithography. Annu. Rev. Mater. Sci. 28, 153–184 (1998). Limnol. Oceanogr. Methods 5, 250–262 (2007).
37. J. G. Cham, S. A. Bailey, J. E. Clark, R. J. Full, M. R. Cutkosky, Fast and robust: Hexapedal 64. X. Tan, D. Kim, N. Usher, D. Laboy, J. Jackson, A. Kapetanovic, J. Rapai, B. Sabadus,
robots via shape deposition manufacturing. Int. J. Rob. Res. 21, 869–882 (2002). X. Zhou, An autonomous robotic fish for mobile sensing, in 2006 IEEE/RSJ International
38. R. Deimel, O. Brock, A compliant hand based on a novel pneumatic actuator, in 2013 IEEE Conference on Intelligent Robots and Systems (IEEE 2006), pp. 5424–5429.
International Conference on Robotics and Automation (ICRA) (IEEE, 2013), pp. 2047–2053. 65. C. Georgiades, A. German, A. Hogue, H. Liu, C. Prahacs, A. Ripsman, R. Sim, L.-A. Torres,
39. A. D. Marchese, R. K. Katzschmann, D. Rus, A recipe for soft fluidic elastomer robots. P. Zhang, M. Buehler, G. Dudek, M. Jenkin, E. Milios, AQUA: An aquatic walking robot,
Soft Robot. 2, 7–25 (2015). in 2004 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (IEEE,
40. R. MacCurdy, R. Katzschmann, Y. Kim, D. Rus, Printable hydraulics: A method for 2004), vol. 4, pp. 3525–3531.

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

fabricating robots by 3D co-printing solids and liquids, in 2016 IEEE International 66. I. Vasilescu, C. Detweiler, M. Doniec, D. Gurdan, S. Sosnowski, J. Stumpf, D. Rus, AMOUR V:
Conference on Robotics and Automation (ICRA) (IEEE, 2016), pp. 3878–3885. A hovering energy efficient underwater robot capable of dynamic payloads. Int. J.
41. R. MacCurdy, J. Lipton, S. Li, D. Rus, Printable programmable viscoelastic materials for Rob. Res. 29, 547–570 (2010).
robots, in 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 67. S. B. Behbahani, X. Tan, Design and modeling of flexible passive rowing joint for robotic
(IEEE, 2016), pp. 2628–2635. fish pectoral fins. IEEE Trans. Robot. 32, 1119–1132 (2016).
42. M. Wehner, M. T. Tolley, Y. Mengüç, Y.-L. Park, A. Mozeika, Y. Ding, C. Onal, R. F. Shepherd, 68. Y. Zhao, M. Fukuhara, T. Usami, Y. Takada, Performance of very small robotic fish
G. M. Whitesides, R. J. Wood, Pneumatic energy sources for autonomous and wearable equipped with CMOS camera. Robotics 4, 421–434 (2015).
soft robotics. Soft Robot. 1, 263–274 (2014). 69. A. Mazumdar, P. Valdivia y Alvarado, K. Youcef-Toumi, Maneuverability of a robotic tuna
43. D. McFarland, I. Gilhespy, E. Honary, DIVEBOT: A diving robot with a whale-like buoyancy with compliant body, in IEEE International Conference on Robotics and Automation 2008
mechanism. Robotica 21, 385–398 (2003). (ICRA 2008) (IEEE, 2008), pp. 683–688.
44. K. Shibuya, Y. Kado, S. Honda, T. Iwamoto, K. Tsutsumi, Underwater robot with a 70. P. Valdivia y Alvarado, K. Youcef-Toumi, Design of machines with compliant bodies for
buoyancy control system based on the spermaceti oil hypothesis, in 2006 IEEE/RSJ biomimetic locomotion in liquid environments. J. Dyn. Syst. Meas. Control. 128, 3–13
International Conference on Intelligent Robots and Systems (IEEE, 2006), pp. 3012–3017. (2005).
45. J. Laine, S. A. Nichols, D. K. Novick, P. D. O’Malley, D. Copeland, A. A. Arroyo, M. C. Nechyba, 71. S. L. Katz, D. A. Syme, R. E. Shadwick, High-speed swimming: Enhanced power in
SubjuGator: A Highly Maneuverable, Intelligent Underwater Vehicle (2015); www.mil.ufl. yellowfin tuna. Nature 410, 770–771 (2001).
edu/publications/fcrar00/subjugator.pdf. 72. J. J. Videler, Fish Swimming (Springer Science & Business Media, 1993).
46. S. Guo, K. Sugimoto, S. Hata, J. Su, K. Oguro, A new type of underwater fish-like 73. A. Maldonado-Ramírez, L. A. Torres-Méndez, F. Rodríguez-Telles, Ethologically
microrobot, in 2000 IEEE/RSJ International Conference on Intelligent Robots and Systems inspired reactive exploration of coral reefs with collision avoidance: Bridging the gap
(IEEE, 2000), vol. 2, pp. 867–872. between human and robot spatial understanding of unstructured environments,
47. R. E. Davis, C. C. Eriksen, C. P. Jones, Autonomous buoyancy-driven underwater vehicle in 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
gliders, in Technology and Applications of Autonomous Underwater Vehicles, G. Griffiths, (IEEE, 2015), pp. 4872–4879.
Ed. (Taylor & Francis, 2002), pp. 37–62. 74. J. Halloy, G. Sempo, G. Caprari, C. Rivault, M. Asadpour, F. Tâche, I. Saïd, V. Durier,
48. C. Detweiler, S. Sosnowski, I. Vasilescu, D. Rus, Saving energy with buoyancy and balance S. Canonge, J. M. Amé, C. Detrain, N. Correll, A. Martinoli, F. Mondada, R. Siegwart,
control for underwater robots with dynamic payloads, in Experimental Robotics, O. Khatib, J. L. Deneubourg, Social integration of robots into groups of cockroaches to control
V. Kumar, G. J. Pappas, Eds. (Springer, 2009), vol. 54 of the Springer Tracts in Advanced self-organized choices. Science 318, 1155–1158 (2007).
Robotics, pp. 429–438. 75. Z. Butler, P. Corke, R. Peterson, D. Rus, From robots to animals: Virtual fences for
49. A. Cloitre, V. Subramaniam, N. Patrikalakis, P. Valdivia y. Alvarado, Design and control of a controlling cattle. Int. J. Rob. Res. 25, 485–508 (2006).
field deployable Batoid Robot, in 2012 4th IEEE RAS & EMBS International Conference on 76. J.-H. Kim, S.-H. Choi, D. Kim, J. Kim, M. Cho, Animal-robot interaction for pet caring, in
Biomedical Robotics and Biomechatronics (BioRob (IEEE, 2012), pp. 707–712. 2009 IEEE International Symposium on Computational Intelligence in Robotics and
50. L. Liu, S. Zhou, J.-H. Cui, Prospects and problems of wireless communication for underwater Automation (CIRA) (IEEE, 2009), pp. 159–164.
sensor networks. Proc. Int. Wirel. Commun. Mob. Comput. Conf. 8, 977–994 (2008). 77. T. Landgraf, R. Rojas, H. Nguyen, F. Kriegel, K. Stettin, Analysis of the waggle dance motion
51. X. Che, I. Wells, G. Dickers, P. Kear, X. Gong, Re-evaluation of RF electromagnetic of honeybees for the design of a biomimetic honeybee robot. PLOS ONE 6, e21354
communication in underwater sensor networks. IEEE Commun. Mag. 48, 143–151 (2010). (2011).
52. I. Vasilescu, K. Kotay, D. Rus, M. Dunbabin, P. Corke, Data collection, storage, and retrieval 78. D. Romano, G. Benelli, C. Stefanini, Escape and surveillance asymmetries in locusts
with an underwater sensor network, in Proceedings of the 3rd International Conference exposed to a Guinea fowl-mimicking robot predator. Sci. Rep. 7, 12825 (2017).
on Embedded Networked Sensor Systems (SenSys 2005) (ACM, 2005), pp. 154–165. 79. N. Abaid, T. Bartolini, S. Macrì, M. Porfiri, Zebrafish responds differentially to a robotic fish
53. M. Doniec, M. Angermann, D. Rus, An end-to-end signal strength model for underwater of varying aspect ratio, tail beat frequency, noise, and color. Behav. Brain Res. 233,
optical communications. IEEE J. Oceanic Eng. 38, 743–757 (2013). 545–553 (2012).
54. M. Doniec, I. Topor, M. Chitre, D. Rus, Autonomous, localization-free underwater data 80. F. Bonnet, Y. Kato, J. Halloy, F. Mondada, Infiltrating the zebrafish swarm: Design,
muling using acoustic and optical communication, in Experimental Robotics, J. P. Desai, implementation and experimental tests of a miniature robotic fish lure for fish-robot
G. Dudek, O. Khatib, V. Kumar, Eds. (Springer, 2013), pp. 841–857. interaction studies, in SWARM 2015: The First International Symposium on Swarm Behavior
55. M. Chitre, S. Shahabudeen, M. Stojanovic, Underwater acoustic communications and and Bio-Inspired Robotics (2016), vol. 21, pp. 239–246.
networking: Recent advances and future challenges. Mar. Technol. Soc. J. 42, 103–116 81. P. Phamduy, G. Polverino, R. C. Fuller, M. Porfiri, Fish and robot dancing together:
(2008). Bluefin killifish females respond differently to the courtship of a robot with varying
56. M. Stojanovic, J. Preisig, Underwater acoustic communication channels: Propagation color morphs. Bioinspir. Biomim. 9, 036021 (2014).
models and statistical characterization. IEEE Commun. Mag. 47, 84–89 (2009). 82. S. Kalantar, U. R. Zimmer, Distributed shape control of homogeneous swarms of
57. P. Lacovara, High-bandwidth underwater communications. Mar. Technol. Soc. J. 42, autonomous underwater vehicles. Auton. Robots. 22, 37–53 (2007).
93–102 (2008). 83. T. Schmickl, R. Thenius, C. Moslinger, J. Timmis, A. Tyrrell, M. Read, J. Hilder, J. Halloy,
58. M. Johnson, D. Herold, J. Catipovic, The design and performance of a compact A. Campo, C. Stefanini, L. Manfredi, S. Orofino, S. Kernbach, T. Dipper, D. Sutantyo,
underwater acoustic network node, in Proceedings of the Oceans Engineering for Today’s CoCoRo—The self-aware underwater swarm, in 2011 Fifth IEEE Conference on Self-Adaptive
Technology and Tomorrow’s Preservation (OCEANS’94) (IEEE, 1994), vol. 3, pp. III–467. and Self-Organizing Systems Workshops (SASOW) (IEEE, 2011), pp. 120–126.

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 11 of 12

SCIENCE ROBOTICS | RESEARCH ARTICLE

84. M. Read, C. Möslinger, T. Dipper, D. Kengyel, J. Hilder, R. Thenius, A. Tyrrell, J. Timmis, 95. L. Amundsen, M. Landrø, Marine seismic sources part VIII: Fish hear a great deal. GEO ExPro
T. Schmickl, Profiling underwater swarm robotic shoaling performance using simulation, 8, 42–46 (2011).
in Towards Autonomous Robotic Systems (Springer, 2013), Lecture Notes in Computer 96. A. N. Popper, D. T. T. Plachta, D. A. Mann, D. Higgs, Response of clupeid fish to ultrasound:
Science, pp. 404–416. A review. ICES J. Mar. Sci. 61, 1057–1061 (2004).
85. J. S. Jaffe, P. J. S. Franks, P. L. D. Roberts, D. Mirza, C. Schurgers, R. Kastner, A. Boch, 97. G. Goertzel, An algorithm for the evaluation of finite trigonometric series. Am. Math. Mon.
A swarm of autonomous miniature underwater robot drifters for exploring submesoscale 65, 34–35 (1958).
ocean dynamics. Nat. Commun. 8, 14189 (2017).
86. N. Correll, Ç. D. Önal, H. Liang, E. Schoenfeld, D. Rus, Soft autonomous materials—Using Acknowledgments: We thank D. Dorhout, A. DeMaille, S. Moon, and J. Wright for their
active elasticity and embedded distributed computation, in 2012 Experimental Robotics: contributions. We are grateful to the reviewers for providing very valuable feedback on earlier
The 12th International Symposium on Experimental Robotics, O. Khatib, V. Kumar, versions of the manuscript. Funding: This work was supported by the NSF via grant numbers NSF
G. Sukhatme, Eds. (Springer, 2014), vol. 79 of Springer Tracts in Advanced Robotics, 1117178, NSF IIS1226883, and NSF CCF1138967 and the NSF Graduate Research Fellowship
pp. 227–240. 1122374. Author contributions: R.K.K. and D.R. conceptualized the robotic fish. R.K.K. developed
87. C. D. Onal, D. Rus, Autonomous undulatory serpentine locomotion utilizing body the robotic fish including design, fabrication, and control, as well as experiments. R.K.K., J.D., R.M.,
dynamics of a fluidic soft robot. Bioinspir. Biomim. 8, 026003 (2013). and D.R. performed the ocean experiments and wrote the paper. R.K.K. and J.D. developed the
88. G. V. Lauder, B. Flammang, S. Alben, Passive robotic models of propulsion by the bodies control software of the robotic fish. J.D. developed the acoustic communication protocol,
and caudal fins of fish. Integr. Comp. Biol. 52, 576–587 (2012). demodulation algorithms, software related to acoustic signal processing and decoding, and
89. S. Alben, C. Witt, T. V. Baker, E. Anderson, G. V. Lauder, Dynamics of freely swimming acoustic experiments. J.D. and R.M. developed the software related to the acoustic transmitter. R.M.
flexible foils. Phys. Fluids 24, 051901 (2012). developed the diver interface module and designed the transmitter and receiver electronics. D.R.
90. R. K. Katzschmann, A. de Maille, D. L. Dorhout, D. Rus, Cyclic hydraulic actuation for soft was responsible for the overall research direction, objectives, and funding. Competing interests:
robotic devices, in 2016 IEEE/RSJ International Conference on Intelligent Robots and The authors declare that they have no competing financial interests. Data and materials
Systems (IROS) (IEEE, 2016), pp. 3048–3055. availability: Contact R.K.K. for source code and other materials.

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

91. K. J. Åström, T. Hägglund, Revisiting the Ziegler–Nichols step response method for PID
control. J. Process Control 14, 635–650 (2004). Submitted 2 November 2017
92. E. J. Tucholski, S. Traffic, Underwater Acoustics and Sonar SP411 Handouts and Notes Accepted 27 February 2018
Fall 2006 (Physics Department, US Naval Academy, 2006). Published 21 March 2018
93. D. H. Cato, Features of ambient noise in shallow water, in Shallow Water Acoustics, 10.1126/scirobotics.aar3449
R. Zhang, J. Zhou, Eds. (China Ocean Press, 1997), pp. 385–390.
94. D. H. Cato, Marine biological choruses observed in tropical waters near Australia. Citation: R. K. Katzschmann, J. DelPreto, R. MacCurdy, D. Rus, Exploration of underwater life
J. Acoust. Soc. Am. 64, 736–743 (1978). with an acoustically controlled soft robotic fish. Sci. Robot. 3, eaar3449 (2018).

Katzschmann et al., Sci. Robot. 3, eaar3449 (2018) 21 March 2018 12 of 12

Exploration of underwater life with an acoustically controlled soft robotic fish
Robert K. Katzschmann, Joseph DelPreto, Robert MacCurdy and Daniela Rus

Sci. Robotics 3, eaar3449.

DOI: 10.1126/scirobotics.aar3449

Downloaded from http://robotics.sciencemag.org/ by guest on March 21, 2018

ARTICLE TOOLS http://robotics.sciencemag.org/content/3/16/eaar3449

SUPPLEMENTARY http://robotics.sciencemag.org/content/suppl/2018/03/19/3.16.eaar3449.DC1
MATERIALS

REFERENCES This article cites 60 articles, 5 of which you can access for free
http://robotics.sciencemag.org/content/3/16/eaar3449#BIBL

PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions

Use of this article is subject to the Terms of Service

Science Robotics (ISSN 2470-9476) is published by the American Association for the Advancement of Science, 1200
New York Avenue NW, Washington, DC 20005. 2017 © The Authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim to original U.S. Government Works. The title
Science Robotics is a registered trademark of AAAS.