Do Rigid-Body Simulators Dream of Soft Robots?
Learning Contact-Rich Manipulation for
Tendon-Driven Continuum Robots

Chengnan Shentu, Nicholas Baldassini, Tongjia Zheng, Priyanka Rao, Jessica Burgner-Kahrs

Continuum Robotics Lab · University of Toronto

The right discretization lets continuum robots live natively inside MuJoCo.

Interactive Demo

Pick a configuration, then click the view and use the keys shown. This browser demo runs the MuJoCo model and teleoperation stack live using WebAssembly and Pyodide.

Initializing…

Abstract

Learning contact-rich, whole-body manipulation for soft continuum robots is held back by the lack of simulation infrastructure that has accelerated rigid-robot manipulation. Existing soft robot simulators are physically grounded but lack the contact handling, actuation support, or learning integration needed for contact-rich manipulation; rigid-body approximations offer these capabilities but sacrifice physical grounding. We bridge this gap for tendon-driven continuum robots (TDCRs) by deriving a continuum-mechanics-informed discretization that places the soft robot natively inside MuJoCo, unifying tendon forces, body contact, and dynamics in a single physics pipeline. We validate the simulator against a Cosserat rod reference (static and dynamic) and real TDCR hardware. We then train state-based imitation learning policies via teleoperation in simulation and deploy them zero-shot to a physical 3-segment TDCR on a 7-DoF Franka arm across two contact-rich manipulation tasks. To our knowledge, this is the first demonstration of sim-to-real transfer for contact-rich manipulation with continuum robots.

TDCR simulation, policy learning, and physical robot deployment overview — Overview of the TDCR-in-MuJoCo pipeline, from mechanics-informed discretization through simulation training and zero-shot deployment.

Contributions

Method. We derive a continuum-mechanics-informed rigid-link discretization that places TDCRs directly in MuJoCo while preserving actuation, contact, and dynamics in one physics pipeline, validated against Cosserat rod simulator and hardware.
System. We build an open workflow around this discretization, including model generation, controllers, teleoperation, and zero-shot sim-to-real policy transfer for contact-rich manipulation.

Mechanics-informed TDCR discretization and learning pipeline overview

Our continuum-informed discretization enables contact-rich zero-shot sim-to-real.

We collect demonstrations in simulation, train state-based imitation-learning policies, and deploy them zero-shot to the 3-segment TDCR on a Franka arm. Both clips are full, uncut rollouts at 1× speed.

Task 1 — Whole-body grasping. The policy must wrap the compliant body evenly around the cylinder, since uneven contact tips it over before a secure grasp.

Task 2 — Flip a switch from behind. The TDCR braces against the panel to localize ~0.9 N on the small toggle, where misalignment or an angled push makes it slip off.

Statics and dynamics stay faithful to a continuum-mechanics reference.

We compare our MuJoCo discretization against SoRoSim, a well-established Cosserat rod solver based on the geometric variable strain approach. The results show that the rigid-link TDCR accurately captures the static equilibria and tip-release dynamics of a continuum-mechanics model.

Static TDCR shapes from MuJoCo compared with SoRoSim references — **Static shapes.** We evaluate on 500 shapes under randomized loading: gravity, midpoint wrenches, and tip wrenches, on two representative materials.

Static convergence errors versus discretization resolution — **Static convergence.** Shape error decreases as the MuJoCo discretization is refined. The two materials track closely despite spanning three orders of magnitude in mechanical stiffness.

Dynamic tip-release trajectories from MuJoCo compared with SoRoSim references — **Tip-release dynamics.** SoRoSim's oscillation frequency and damping decay are recovered exactly, across simulation rates.