Soft robotics · Underwater locomotion

A Tendon-Driven Robotic Jellyfish
with Constrained Soft Actuation and
Depth Control via Reinforcement Learning

Jiarui Peng1,2,Yutong Wu1,2,Zelong Wang3,Ping Deng3,Xiaotian Zhang3,Xian Chen3,Kecheng Qin1,2,*,Zhongyu Li1,2,*
1Hong Kong Embodied AI Lab2The Chinese University of Hong Kong3The Hong Kong University of Science and Technology

*Corresponding authors

Hong Kong Centre for Logistics Robotics
Hong Kong Embodied AI Lab The Chinese University of Hong Kong The Hong Kong University of Science and Technology
Paper arXiv Video Code Coming Soon
A cross jelly (Mitrocoma cellularia) beside the tendon-driven robot, whose translucent silicone bell is supported by eight radial actuators.
Left: cross jelly (Mitrocoma cellularia). Right: JellyRobot.Inspired by nature. Constrained by design. Controlled through learning.

A compliant jellyfish robot that pairs predictable tendon-driven bending with a learned, cycle-level controller for underwater depth regulation.

150°
Actuator bending range
8 actuators · 4 servos
One compliant swimming platform
Sim-to-real
Depth control without policy fine-tuning

See it in motion

Overview Video

Swimming, manoeuvring, and learned depth regulation.

The idea

Abstract

Jellyfish-inspired robots offer a compliant and efficient approach to underwater locomotion, but achieving large deformation together with repeatable actuation and closed-loop control remains challenging. In this work, we present a tendon-driven robotic jellyfish with constrained soft actuation. Each actuator combines a flexible substrate with discrete constraints, enabling bending up to 150° with an approximately linear tendon displacement–bending relationship.

Eight actuators driven by four servos allow the robot to perform stable swimming, attitude adjustment, and self-righting. Based on the linear actuation, a reinforcement-learning controller is further developed, enabling closed-loop depth regulation in both simulation and physical experiments. These results show that mechanical constraints can improve the controllability of soft actuation while preserving compliant jellyfish-like motion, providing a route towards manoeuvrable and autonomous jellyfish robots.

01 / Mechanical design

Soft actuation. Predictable bending.

A flexible TPU substrate provides elastic recovery. Discrete PLA constraints guide large, repeatable deformation.

Robot architecture and a layered actuator: TPU substrate, PLA limiting ossicles, tendon, and membrane clips.
Robot architecture. Four servomotors each drive a pair of neighbouring radial actuators beneath a compliant silicone bell.
Approximately linear total bending angle against tendon take-up, with measured configurations up to 150 degrees at 29.72 millimetres.
Actuator characterisation. A tendon take-up of 29.72 mm produces approximately 150° of bending with a quasi-linear displacement–bending relationship.
Tendon force characterisation Air and water comparison
Contraction and release force curves, showing peak tendon loads of 27.6 newtons in water and 24.5 newtons in air.
Contraction and recovery follow different force paths. Shading denotes one standard deviation; these measurements characterise tendon loading, not propulsive thrust.

02 / Swimming behaviours

One bell. Multiple skills.

Synchronous strokes generate ascent. Asymmetric amplitudes enable steering. Passive mechanical trim supports self-righting.

Four servo channels and a prescribed cycle progressing from relaxed through contraction, maximum bending, release, and recovery.
A prescribed contraction–recovery gait keeps the four channels in phase while allowing independent stroke amplitudes.
Tank experiments showing synchronous ascent, asymmetric steering, and recovery from inversion, alongside displacement and attitude measurements.
Physical swimming experiments. Ascent, steering, and recovery from inversion under prescribed open-loop actuation. Self-righting occurs during continued pulsation without learned attitude control.

03 / Learning to regulate depth

One decision per swimming cycle.

The PPO policy adjusts a single common stroke amplitude from depth error, vertical velocity, and the previous action.

Cycle-synchronous depth control pipeline: sensor feedback, policy, prescribed stroke, and robot, with simulation rollouts, reward, and PPO updates used only during training.
Train in simulation. Deploy the actor on the robot. The selected amplitude is held for the complete stroke. The observation interface, action mapping, and low-level timing are preserved during physical deployment.
01

Sense

Estimate the vertical state from depth and inertial feedback.

02

Adjust

Select a shared amplitude for all four actuation channels.

03

Swim

Complete the prescribed stroke, then evaluate the policy again.

04 / Experimental evaluation

From simulation to the water.

Learned amplitude adjustment improves simulated depth tracking and transfers directly to the physical robot.

8.31 → 2.83 cm
Aggregate simulated depth RMSE
Fixed amplitude → PPO
≈66%
Reduction in depth RMSE
Across 16 simulated initial conditions
5.3 cm
Physical depth RMSE
Representative deployment trial
Simulated and physical swimming sequences with depth traces. The illustrated episodes have RMSEs of 2.52 centimetres in simulation and 5.3 centimetres on the robot.
Representative closed-loop depth regulation. The figure’s annotated RMSEs describe the illustrated trajectories, rather than multi-trial aggregates. Physical deployment uses the trained policy without additional policy training or fine-tuning.

Physical tests used a glass tank with slack external connections for power and communication. Passive mechanical trim remains the principal attitude-restoring mechanism.

Acknowledgements

This work was in part supported by The Chinese University of Hong Kong, the InnoHK initiative of the Innovation and Technology Commission of the Hong Kong Special Administrative Region Government via the Hong Kong Centre for Logistics Robotics, The Hong Kong University of Science and Technology, and the National Natural Science Foundation of China.

Hong Kong Centre for Logistics Robotics Hong Kong Embodied AI Lab The Chinese University of Hong Kong The Hong Kong University of Science and Technology

Scroll to inspect · Press Escape to close