💻 computer science

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

This paper bridges the gap between physically grounded soft robot simulation and learning-based control by integrating tendon-driven continuum robots natively into MuJoCo, enabling the first zero-shot sim-to-real transfer of contact-rich manipulation policies from simulation to physical hardware.

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

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a very flexible, noodle-like robot arm how to do delicate tasks, like picking up a slippery egg or flipping a light switch from behind a wall. In the real world, this is incredibly hard. The robot is floppy, it has many moving parts, and if you try to teach it by trial and error on the actual machine, you might break it, or it might take years to learn.

Usually, engineers solve this by teaching the robot in a computer simulation first. But here's the catch: most simulators are built for rigid robots (like a human arm made of metal bones). They are great at handling hard objects hitting other hard objects, but they are terrible at simulating soft, squishy things that bend and twist.

On the other hand, simulators built for soft robots are physically accurate but are often too slow, clunky, or missing the features needed to learn complex tasks like "grabbing" or "pushing."

This paper introduces a clever new way to bridge that gap. Here is the breakdown of what they did, using simple analogies:

1. The "Lego" Trick (The Core Innovation)

The researchers wanted to put a soft, tendon-driven robot (a robot pulled by strings, like a puppet) inside MuJoCo, a super-fast, industry-standard physics engine designed for rigid robots.

Normally, you can't just drop a soft noodle into a rigid engine; the engine doesn't know how to handle it. So, the authors invented a new way to describe the soft robot.

The Metaphor: Imagine the soft robot is a long, flexible snake. Instead of trying to simulate the whole snake as one continuous, squishy blob, they broke it down into a chain of 30 tiny, stiff Lego blocks connected by tiny, stretchy rubber bands.
The Magic: They didn't just guess how stiff those rubber bands should be. They used advanced math (continuum mechanics) to calculate the exact stiffness based on the real material properties of the robot.
The Result: To the MuJoCo engine, the robot looks like a chain of rigid blocks. But because the "rubber bands" are calculated perfectly, the chain bends and twists exactly like the real soft robot would. This allows the engine to use its super-fast "contact solver" (the part that figures out how things bump into each other) on a soft robot for the first time.

2. The "Teleporting" Teacher (Sim-to-Real)

Once they had this perfect simulation, they used a standard learning workflow:

Teleoperation: A human played with the robot in the simulation using a keyboard, showing it how to do tasks.
Learning: An AI watched these demonstrations and learned the rules of how to move the robot's "strings" (tendons) to achieve the goal.
Zero-Shot Transfer: This is the big claim. They took the AI brain they trained in the computer and instantly put it onto the real physical robot. They did not retrain it on the real machine. They didn't tweak it. They just turned it on.

3. The Proof: Two Tough Tasks

They tested this "teleporting" method on two very difficult tasks that require the robot to use its whole body to touch and push things:

Task A: The "Hug" (Grasping): The robot had to wrap its soft body around a cylinder and lift it into a bin. It couldn't just use a gripper; it had to use its whole flexible body to "hug" the object securely.
- The Result: The robot succeeded 76% of the time in the real world, which was actually slightly better than in the simulation!
Task B: The "Sneak Attack" (Flipping a Switch): The robot had to reach behind a wall and press a light switch down. This is hard because the robot is floppy; usually, it would just bend uselessly against the wall.
- The Result: The robot learned to brace itself against the wall and push down with just the right amount of force. It succeeded 78% of the time in the real world.

Why This Matters

The authors claim this is the first time a soft, continuum robot has successfully learned a complex, contact-heavy task in simulation and then performed it perfectly on the real hardware without any extra tuning.

They proved that by treating a soft robot as a "smart chain of rigid blocks," you can use the best, fastest tools available for rigid robots to train soft robots. This opens the door to teaching soft robots to do complex jobs much faster and cheaper than before, without needing to break expensive hardware during the learning process.

In short: They figured out how to trick a rigid-robot simulator into thinking a soft robot is rigid, but with such perfect math that the soft robot behaves exactly like the real thing. This allowed them to train the robot in a video game and have it perform perfectly in real life immediately.

Technical Summary: Learning Contact-Rich Manipulation for Tendon-Driven Continuum Robots

Problem Statement

Learning contact-rich, whole-body manipulation for soft continuum robots has been hindered by a lack of simulation infrastructure comparable to that available for rigid robots. While rigid-body manipulation has advanced rapidly due to mature simulators, soft continuum robots face a "simulation gap." Existing soft robot simulators are often physically grounded (based on continuum mechanics) but lack robust contact solvers, native actuation modeling (specifically for tendons), or integration with learning pipelines. Conversely, rigid-body approximations (pseudo-rigid-body models) offer native contact and learning integration but often rely on empirically calibrated parameters that lack physical grounding and convergence guarantees. Consequently, progress in learning contact-rich manipulation for soft robots remains limited, as researchers cannot easily iterate on hardware or generate the extensive demonstration data required for learning.

Methodology

The authors propose a continuum-mechanics-informed discretization that places Tendon-Driven Continuum Robots (TDCRs) natively inside the MuJoCo physics engine without modifying the engine itself. This approach bridges the gap between physical fidelity and simulation efficiency.

1. Continuum-Informed Discretization

The TDCR is modeled as a Kirchhoff rod, a specialization of the Cosserat rod model suitable for thin elastic rods where bending and twisting dominate over shear and extension.

Geometric Construction: The rod is divided into $N$ segments. Each segment is approximated as a circular arc under a piecewise constant strain (PCS) assumption.
Rigid-Body Representation: The discretization places a serial chain of rigid links connected by 3-DOF elastic spherical joints inside MuJoCo.
Parameter Derivation: Unlike prior pseudo-rigid-body models that rely on empirical calibration, the joint stiffnesses ( $k_x, k_y, k_z$ ) are derived analytically from the rod's material properties (Young's modulus $E$ , shear modulus $G$ , and cross-sectional moments of inertia $I, J$ ) and segment length $l_i$ :
$k_x = k_y = \frac{EI}{l_i}, \quad k_z = \frac{GJ}{l_i}$
Convergence: The discretization is proven to converge to the continuous Kirchhoff rod solution at a second-order rate ( $O(1/N^2)$ ) as the number of segments increases.
Native Integration: The model utilizes MuJoCo's native tendon routing and contact solver. Tendons are routed through discrete sites on the rigid links, naturally capturing inter-tendon coupling and tension-induced stiffness without abstracting them as distributed loads.

2. Sim-to-Real Pipeline

The authors establish a complete workflow for contact-rich manipulation:

Model Generation: A MuJoCo model is generated using physical parameters.
System Identification: Uncertain parameters (stiffness, tendon pre-tension) are identified against real hardware using Bayesian optimization to minimize tip position error.
Data Collection: Teleoperated demonstrations are collected in simulation using a state-based interface.
Imitation Learning: Policies are trained using a state-based variant of the ACT (Action Chunking with Transformers) algorithm.
Zero-Shot Deployment: Policies are deployed directly to the physical robot without fine-tuning.

Key Contributions

Simulation Framework: The derivation of a discretization that places TDCRs natively in MuJoCo, unifying tendon forces, body contact, and dynamics in a single physics pipeline. This is the first approach to combine physical grounding (via continuum mechanics), native contact resolution, and learning integration for soft continuum robots.
System Validation: Validation of the simulator against a Cosserat rod reference (SoRoSim) and real hardware. The simulation achieves sub-percent error in statics and dynamics against the reference and a mean tip tracking error of 4.1% against physical hardware after system identification.
Sim-to-Real Transfer: The first demonstration of sim-to-real transfer for contact-rich manipulation with continuum robots. The authors successfully trained policies in simulation and deployed them zero-shot to a physical 3-segment, 9-tendon TDCR mounted on a 7-DoF Franka arm.

Results

The paper evaluates the approach across three levels:

Simulation vs. Theory: The MuJoCo discretization matches the SoRoSim Cosserat reference with high fidelity. Static tip errors are below 0.2% of rod length at $N=50$ , and dynamic errors remain below 1% across various discretization levels.
Simulation vs. Hardware: After system identification, the simulator tracks a physical TDCR with a mean tip error of 7.7 mm (4.1% of the robot's 186.5 mm length) across a held-out test set, despite unmodeled effects like tendon friction and hysteresis.
Task Performance: The authors tested two contact-rich tasks:
1. Whole-body Grasping: Wrapping the compliant body around a cylinder to lift and drop it.
2. Flip Switch: Approaching a wall-mounted switch from behind and flipping it using distributed contact.
Policies trained in simulation achieved success rates of 76% (Grasping) and 78% (Switch) on the real hardware, slightly exceeding the simulation success rates (73% for both). This indicates that the simulator's fidelity is sufficient for transfer and that the distributed compliance of the robot accommodates residual errors.

Significance and Claims

The paper claims to be the first demonstration of sim-to-real transfer for contact-rich manipulation with continuum robots. By deriving a discretization that is both physically grounded and compatible with standard rigid-body engines, the authors enable the application of mature simulation-driven learning paradigms (teleoperation, imitation learning, zero-shot deployment) to soft robotics.

The work suggests that the field of soft continuum robotics can now leverage the same data collection, scaling, and benchmarking workflows that have accelerated rigid-robot manipulation. The authors emphasize that their approach restores physical interpretability to pseudo-rigid-body models, allowing for convergence guarantees and generalization beyond calibration conditions, while unlocking the contact resolution and learning integration previously unavailable to this class of robots.

The paper remains modest regarding its scope, noting that the validation is specific to slender, thin-backbone TDCRs where the Kirchhoff assumption holds, and that the learning algorithm (ACT) was not compared against alternative architectures. Future work is identified as extending the discretization to general Cosserat rods (with shear/extension) and exploring vision-based shape feedback.