Nature Communications · 2026 Accepted

Learning Soft Robotic Dynamics with Active Exploration

SoftAE lets a soft robot collect its own training data by exploring where its learned dynamics model is least certain. The result is one task-agnostic model that plans new tasks zero-shot.

All at ETH Zurich · *Corresponding author

Real data · replay

Where did each method collect its data?

The exploration data released with the paper, replayed: 100 episodes on the simulated soft arm, each run as four parallel simulations. Each arm’s tip lays down the dots, sampled every 0.05 s, and faint lines show its earlier poses. Positions are projected and cut at the ground as in the paper’s Fig. 3A. Source: replay buffers on Zenodo.

Supplementary Movie 1

From exploration to zero-shot control.

The overview film walks through the method and all four robots.

Play the overview video Loads YouTube
Six tracking results drawn in red, one letter each, spelling SOFTAE: the soft arm, the musculoskeletal leg, the real SoPrA arm and the fish each trace part of the word.
Fig. 1F Downstream trajectory tracking on the four platforms spells the method’s name.

01 The idea

The data decides what a model can do.

Soft bodies deform continuously, so their dynamics are high-dimensional, nonlinear and hard to model from first principles. A learned model is only as general as its data, and how a soft robot gathers that data matters.

  1. Random exploration: the data spreads only a little way from the robot. On a new task, the robot reaches the red goal but misses the green one.

    Random exploration

    Independent random actions largely cancel out on a compliant body, so data piles up near the rest pose. Tasks that need large deformations stay out of reach.

  2. Task-oriented exploration: data follows one narrow path to the red goal. A new green goal fails.

    Task-oriented learning

    Collecting data along one task’s trajectory gives a model that serves that task and overfits. A new goal falls outside what it has seen.

  3. Active exploration: data reaches into the uncertain regions around the robot. Both the red and the green goal succeed.

    Active exploration (SoftAE)

    Exploring where the model is uncertain yields diverse, informative data and a task-agnostic model that also solves tasks it never trained for.

Fig. 1E Top row: collecting data. Bottom row: planning a new task with the learned model.

02 How SoftAE works

Uncertainty is the only reward.

No task is given during exploration. SoftAE rewards the robot for visiting states where its own dynamics model is unsure, then learns from what it finds there.

  1. 01

    Learn an ensemble

    Several neural networks are fitted to all data so far. Their mean μ predicts the next state; their disagreement σ measures what the model does not know.

  2. 02

    Plan optimistically

    Model-predictive control with iCEM searches for actions, and for plausible dynamics inside μ ± βσ, that collect the most uncertainty along the way.

  3. 03

    Explore on the robot

    The robot runs the plan and records what actually happens, in the regions its model knows least about.

  4. 04

    Update and repeat

    New data shrinks the uncertainty there, so the next episode is drawn somewhere new.

Then, zero-shot

Swap the exploration reward for any task reward r(s, a). The same model and the same planner solve the task, with no retraining and no fine-tuning.

Why optimistic planning helps, and the full exploration objective, are in the paper’s Methods .

03 Zero-shot results

Four soft robots. Ten tasks. One recipe.

SoftAE and Random never see a task while exploring, so every task below is planned zero-shot. The task-specific baseline H-UCRL trains on tasks (i), (iii), (v), (vii) and (ix); the others are unseen for it too.

RandomH-UCRL, task-specificSoftAE† H-UCRL training task
Supplementary Movie 2 Setup, tasks (i)–(ii) and how returns grow with more SoftAE exploration

Reaching the far target (ii) takes extreme bending. The H-UCRL model, trained on the close target, does not generalise there. SoftAE learns it fastest and reaches the highest return.

Supplementary Movie 3 Fluid simulation, fluid–structure interaction and swimming tasks (iii)–(vi)

Swimming needs coordinated spine undulation against delayed hydrodynamic feedback. Random rarely produces it and fails on (v)–(vi). SoftAE even beats H-UCRL on (v), a task H-UCRL was trained on.

Supplementary Movie 4 Electrohydraulic muscle actuation, setup and hopping tasks (vii)–(viii)

Hopping needs precise muscle–motor coordination and ground-contact timing, which random exploration does not discover. SoftAE hops in both directions and beats H-UCRL even on (vii), its training task.

Supplementary Movie 5 Real-world setup and task performance, Fig. 4 in motion

On hardware, sensing noise, actuation delays, hysteresis and pneumatic variability all come into play, and exploration and planning stay the same. Near the edge of the workspace SoftAE outperforms both baselines, and across 20 random 3-D targets its mean tip error is 2.00 mm.

Model-free is not an option at this budget. Goal-conditioned SAC trained directly on the arm made no progress after 40,000 real transitions, about 12 hours and twice the data the model-based methods used (Supplementary Fig. 5).

04 Stress tests

More dimensions. Hard contact.

The same exploration loop holds up when the state gets much larger, and when the soft arm has to work around rigid obstacles.

Fig. 3C

136-dimensional state

Tracking 11 points along the arm instead of 5 raises the state from 58 to 136 dimensions. SoftAE keeps learning quickly to a high return; Random degrades slightly.

Fig. 3D

Learning soft–rigid contact

A model trained without obstacles predicts the arm passing straight through one. Trained among obstacles, it reproduces the contact-induced deformation over a 10-step rollout.

Fig. 3E

Reaching around obstacles

Reaching under a bar, or through a gap at the base of a wall, needs coordinated deformation. Only SoftAE consistently succeeds at both.

Beyond model-predictive control. With policies trained on model rollouts (an MBPO variant using SAC), SoftAE matches task-specific MBPO on its training task and outperforms it on unseen ones (Supplementary Fig. 4).

05 Paper & code

Read it. Run it.

Learning Soft Robotic Dynamics with Active Exploration

Hehui Zheng, Bhavya Sukhija, Chenhao Li, Klemens Iten, Andreas Krause, Robert K. Katzschmann

Nature Communications (2026), accepted, in press · Preprint arXiv:2510.27428

Abstract

Soft robots offer adaptability and safety in unstructured environments, yet their compliant, high-dimensional, and nonlinear dynamics make modeling for control difficult. Existing data-driven approaches often fail to generalize, constrained by narrowly focused task demonstrations or inefficient random exploration. We introduce an uncertainty-aware active exploration framework that autonomously learns task-agnostic and generalizable dynamics models of soft robotic systems. We evaluate this framework on three simulated platforms—a continuum arm, an articulated fish in fluid, and a musculoskeletal leg—and on a real-world pneumatically actuated continuum arm. Compared with random exploration and task-specific baseline, our method achieves higher modeling accuracy by actively collecting informative data. This improved model quality yields higher returns in zero-shot control on unseen tasks and robust performance under sensing noise, actuation delays, and material nonlinearities. These results demonstrate that uncertainty-driven active exploration enables data-efficient dynamics learning across soft robotic morphologies, supporting autonomous and adaptable control for compliant robots.

Citation

@misc{zheng2025learningsoftroboticdynamics,
  title         = {Learning Soft Robotic Dynamics with Active Exploration},
  author        = {Hehui Zheng and Bhavya Sukhija and Chenhao Li and
                   Klemens Iten and Andreas Krause and Robert K. Katzschmann},
  year          = {2025},
  eprint        = {2510.27428},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2510.27428}
}

Questions about the work?

Enlarged figure

Prediction error on held-out data

Normalised mean squared error of next-state predictions over 17,451 transitions · lower is better

Normalised mean squared error of learned dynamics models, Fig. 3B
MethodNormalised MSE
Random0.2397
H-UCRL0.7232
SoftAE0.0526

4.6× lower than Random · 14× lower than H-UCRL

Fig. 3B Dynamics models trained on each method’s exploration data from the simulated soft arm, tested on rollouts towards 500 targets sampled uniformly across the workspace.

Tip error over 20 random 3-D targets

mean best to worst target · millimetres, 3 seeds

Random: mean 3.82, worst 9.57 millimetres. H-UCRL: mean 10.15, worst 49.41. SoftAE: mean 2.00, worst 4.70. 01020304050 Random 3.829.57 H-UCRL 10.1549.41 SoftAE 2.004.70
Real-world reaching over 20 random targets, Fig. 4D
MethodMean returnMean tip error (mm)Std (mm)Min (mm)Max (mm)
Random17.633.822.001.209.57
H-UCRL15.2310.1513.711.1449.41
SoftAE17.782.001.130.494.70

Worst case 4.70 mm for SoftAE against 49.41 mm for H-UCRL, more than ten times lower.

Fig. 4D The real SoPrA arm, after the same exploration budget for each method. SoftAE also has the highest mean return (17.78; Random 17.63, H-UCRL 15.23).

Ten downstream tasks, all planned zero-shot

SoftAE and Random see no task while exploring. † marks the tasks the task-specific baseline H-UCRL trains on.

PlatformTasksState · action
Simulation
  • (i) Reach a close target†
  • (ii) Reach a far target
58-D · 12-D
Simulation
  • (iii) Swim along +x†
  • (iv) Turn and swim along −x
  • (v) Swim to a target†
  • (vi) Swim away from a target
15-D or 18-D · 4-D
Simulation
  • (vii) Hop counter-clockwise†
  • (viii) Hop clockwise
10-D · 5-D
Real hardware
  • (ix) Reach a close target†
  • (x) Reach a far target
32-D · 6-D

Also tested: trajectory tracking on SoPrA, and reaching around obstacles with the soft arm. Pick a platform to jump to its results.