All You Need Is Low Fidelity: Zero-Shot Sim-to-Real of Learned Robotic Fish Control

Liam Maloney; Simon Ramchandani; Mike Y. Michelis; Ronan Hinchet; Robert K. Katzschmann

cs.ROSep 29, 2026

All You Need Is Low Fidelity: Zero-Shot Sim-to-Real of Learned Robotic Fish Control

Authors: Liam Maloney, Simon Ramchandani, Mike Y. Michelis, Ronan Hinchet, Robert K. Katzschmann

Organizations: Soft Robotics Lab, ETH Zurich, Switzerland · ETH AI Center, ETH Zurich, Switzerland

Abstract

Complex tasks for underwater robots remain limited by the capabilities of their controllers. Learning a better one for a soft, underactuated robotic fish trades simulator cost against fidelity. We show that an intentionally low-fidelity simulator is enough: a stateless, quasi-steady fluid model with no wake and no added-mass history suffices to learn a \emph{general}, closed-loop controller that transfers to hardware without tuning. Our platform is a soft, single-motor, tendon-driven fish whose policy observes only what the hardware can measure. A staged pipeline grounds the simulator in two independent identifications, fixing the tail dynamics and a stateless fluid model; the policy then acts through a band-limited rhythmic trajectory generator rather than commanding the tail directly. Deployed unchanged in an outdoor pool, a single policy performs closed-loop target reaching, disturbance rejection, and out-of-distribution target acquisition and tracking. The transfer rests on the constraint rather than the fidelity: the generator cannot leave the band over which the fluid was identified. This raises the question of how much of the physics can reside in the controller rather than in the simulator.

Figures & tables

Fig. 1 : Overview. Separate experiments identify the tail dynamics and a stateless fluid model. The policy sets the parameters $\mathbf{u}^{*}$ of a band-limited trajectory generator (TG) that emits the motor setpoint $\theta_{m}$ . The same policy and TG run unchanged onboard the real soft fish robot, closed by hardware observations.

Fig. 2 : Policy training loop. Left: the MuJoCo twin and the fifteen quasi-steady fluid coefficients identified for its three zones ( Section IV-B , Table I ). Center: the policy commands target gait parameters $\mathbf{u}^{*}$ rather than motor angles. These are smoothed by a first-order time lag $\tau$ into $\mathbf{u}$ , which the trajectory generator converts into motor setpoint $\theta_{m}$ ( Equation 5 ). The observed state holds only what the hardware can measure. Right: soft actor-critic, rewarded for closing the distance to the target ( Equation 7 ).

Fig. 3 : A: Translucent CAD model of the hardware fish. A sealed, oil-filled head carries the onboard compute; the motor in the flooded body turns the pulley that drives the antagonistic tendon pair along the compliant tail. B: The assembled fish as deployed, with the AprilTag on the head that the overhead camera tracks.

Zone	blunt	slender	angular	Kutta	Magnus
Head/Body	$0.006$	$0.045$	$0.813$	$0.521$	$0.903$
Tail	$0.039$	$0.551$	$5.428$	$0.001$	$1.581$
Fin	$0.171$	$0.004$	$18.310$	$10.148$	$5.951$

TABLE I : Identified quasi-steady fluid coefficients, five per zone. The angular drag and Kutta lift of the fin are markedly larger, the stateless model inflating them to stand in for unsteady thrust.

Parameter	Idealized	Identified
Position gain $k_{p}$	$800$	$5.41$
Velocity gain $k_{v}$	$20$	$0$
Rotor armature [ $\mathrm{kg}\text{$ {}\cdot{} $}{\mathrm{m}}^{2}$ ]	$0.001$	$0.0113$
Joint damping $b$ [ $\mathrm{N}\text{$ {}\cdot{} $}\mathrm{m}\text{$ {}\cdot{} $}\mathrm{s}\text{/}\mathrm{rad}$ ]	$0$	$0.48$
Force limit [ $\mathrm{N}\text{$ {}\cdot{} $}\mathrm{m}$ ]	$1.0$	$2.3$
Transport delay $\tau_{d}$ [ $\mathrm{ms}$ ]	$0$	${\approx}14$

TABLE II : Motor actuator parameters: the idealized model the deployed policy was trained on, against the motor model identified afterwards. Position gain and rotor armature are fitted; the force limit is the hardware value; the joint damping caps peak shaft speed near the measured

5\text{\,}\mathrm{rad}\text{/}\mathrm{s}

Fig. 7 : Compilation of paths swum by the deployed soft fish. A: Overhead pool composites of four fixed-target runs, each panel titled by the condition it tests, with fish snapshots at fixed intervals over the full swum path and the AprilTag-marked target. Notes read initial distance, closest approach and elapsed time. B: Dynamic tracking of a target moved by hand along a figure-eight for about $73\text{\,}\mathrm{s}$ . Top, the overhead composite; bottom, the tracked fish (solid) and target (dotted) paths, markers every $10\text{\,}\mathrm{s}$ .

Explore similar work

FinsSim: A Reality-Aligned Integrated Simulation Platform for Underwater Robot Learning
Sep 20, 2026Yu Zhang, Yuanmingqing Song, Xiangyun Rao +4Underwater Robots
Department of Automation, Shanghai Jiao Tong University, Shanghai, China · School of Ocean and Civil Engineering, Shanghai Jiao Tong University, Shanghai, China · Zhejiang University, Hangzhou, China
Differentiable Reinforcement Learning for Path Tracking by an Agile Fish-Like Robot
Jul 17, 2026Prashanth Chivkula, Kartik Loya, Venkata Ravindhra Reddy Varikuti +1Underactuated Dynamics Motion Planning
Department of Mechanical Engineering, Clemson University, USA
Video2SwimFish: An Automated Pipeline for Reconstructing Controllable Fish Models and Biological Locomotion from Real Fish Videos
Sep 30, 2026Hangong Chen, Linfeng Cheng, Tahsin Zaman Jilan +4Fish Underwater Robots
Department of Computer Science, North Carolina A&T State University · Department of Computer Science, University of Maryland, College Park

Sep 20, 2026, cs.ROJ/K move · Enter open · S save

FinsSim: A Reality-Aligned Integrated Simulation Platform for Underwater Robot Learning

Yu Zhang, Yuanmingqing Song, Xiangyun Rao, Pangkit Fong+3

Department of Automation, Shanghai Jiao Tong University, Shanghai, China · School of Ocean and Civil Engineering, Shanghai Jiao Tong University, Shanghai, China · Zhejiang University, Hangzhou, China

Underwater Robots

Open paper AskPDF arXiv

Underwater robot learning relies on simulators that integrate high-fidelity hydrodynamics, convenient learning interfaces, and a credible transition to real scenarios. In this work, we present FinsSim, a reality-aligned integrated simulation platform for Sim-to-Real underwater robot learning. FinsSim first constructs high-fidelity simulation with selectable backends to adapt to diverse requirements. To facilitate underwater robot research, it further offers standard control baselines, alongside with unified robot learning workflows. For reliable Sim-to-Real transfer, FinsSim adopts a multi-sensor fusion scheme to provide low-cost yet precise localization. Moreover, it implements calibrated thruster-hydrodynamics models and a constrained wrench allocation algorithm. Bridging these modules by ROS~2, FinsSim establishes a complete Sim-to-Real transfer pipeline. Through matched simulations and experiments, it is demonstrated that reliable Sim-to-Real transfer of underwater robot control policies can be achieved with the FinsSim framework. Separate ablation studies also validate that the modules of FinsSim can address the pivotal issues of underwater Sim-to-Real from different aspects. Overall, this work aims to bridge the gap between theoretical research and practical applications, ultimately driving advancements in the field of underwater robotics.

Abstract

Explore similar work

FinsSim: A Reality-Aligned Integrated Simulation Platform for Underwater Robot Learning

Differentiable Reinforcement Learning for Path Tracking by an Agile Fish-Like Robot

Video2SwimFish: An Automated Pipeline for Reconstructing Controllable Fish Models and Biological Locomotion from Real Fish Videos