Should We Learn Contact-Rich Manipulation Policies From Sampling-Based Planners?

Authors: Huaijiang Zhu, Tong Zhao, Xinpei Ni, Jiuguang Wang, Kuan Fang, Ludovic Righetti, Tao Pang · RA-L 2025

Licences

Open Access
Paper: CC BY 4.0

Everything below draws on the full paper under its CC BY 4.0 licence; its figure is credited to the authors.

TL;DR

RRT-generated demonstrations have high action entropy that hurts behavior cloning; more consistent planner data trains diffusion policies for contact-rich tasks that transfer zero-shot to hardware.

Why it matters

The paper shows that the planner's design, not only the amount of data, decides whether behavior cloning from planned demonstrations works: lower-entropy data beat probabilistically complete RRT data even though RRT covered states slightly better. It offers a way to get training data for tasks with full-arm contacts and multi-finger coordination, which current teleoperation interfaces cannot easily demonstrate.

Problem

Behavior cloning has worked mainly where humans can teleoperate demonstrations, but common teleoperation interfaces only track end effectors, so demonstrations with full-arm contact or multi-finger coordination are hard to collect. Reinforcement learning as a data source treats constraints as reward penalties, which complicates reward shaping and sim-to-real transfer. The paper asks how model-based contact planners should generate training data for such tasks.

Proposed solution

A planner using a smoothed, quasi-dynamic contact model proposes a plan, which is re-executed in chunks in a physics simulator without contact smoothing to produce state-action demonstrations; reached states are then relabelled as goals (hindsight goal relabelling). A goal-conditioned diffusion policy (1D U-Net with FiLM conditioning) is trained on this data. The planner is changed for consistency: a greedy search for the bimanual task, and a sparse probabilistic roadmap over the 24 canonical cube orientations with fixed primitives and pre-computed grasps for the in-hand task.

Key insights

RRT samples subgoals at every tree expansion, which yields demonstrations with high action entropy that are hard to learn from, especially with little data.
Better state coverage did not give better policies: RRT data covered slightly more states yet trained clearly worse policies than greedy-planner data.
Hindsight goal relabelling turns sub-optimal or failed demonstrations into valid training pairs for the states they actually reached.
For the cube task, a hybrid of a main policy and a final-adjustment policy counters data imbalance and raised simulated success by about 10% over one unified policy.
Long action horizons (40 and 60 steps) were used because short horizons gave jerky motion from switching between action modes.

Main results

IiwaBimanual in simulation (100 initial positions): greedy-planner data reaches 99% success with 100 demos versus 44% for Contact-RRT data; with 500, 1,000 and 5,000 demos it is 98%, 99% and 100% versus 63%, 88% and 84%.
Simulation, 100 initial poses: AllegroHand-Easy 82% and AllegroHand-Hard 68% with the hybrid policy, versus 74% and 57% for a unified policy; IiwaBimanual 99%.
Hardware AllegroHand (hybrid policy): 15 of 24 trials succeed for each variant (62.5%), with orientation errors on successful trials of 13.7 ± 5.9° (Easy) and 15.5 ± 5.2° (Hard).
Hardware IiwaBimanual: 18 of 20 trials succeed (90%); failed trials ended at about 15° orientation error, just above the 0.2 rad (11.5°) threshold.
Transfer was zero-shot despite model mismatch (real object 1.25 kg and 0.59 m diameter versus 1.0 kg and 0.6 m in simulation) and without domain randomization.

Limitations

The policy often fails to react to out-of-distribution states absent from training data, a consequence of the pre-computed primitives; this happens more often on hardware.
DAgger-style correction is hard to apply because the planner cannot find solutions from arbitrary system configurations.
The IiwaBimanual policy sometimes chatters between clockwise and counter-clockwise rotation, so it often takes longer than necessary; the AllegroHand policy sometimes uses longer primitive sequences than needed.
The best data-mixture ratio for a unified AllegroHand policy is left open, and warm-starting inference with the previous action prediction to reduce jerk is left for future work.
No domain randomization of object or robot geometry or friction was used during training or data generation.

Field impact

The paper won the RA-L Best Paper Award (listed as RA-L 2026) and has 23 citations. It releases a hardware experiment video (https://youtu.be/CxgjJmiiEhI); the paper names no code or dataset release.

Abstract

The tremendous success of behavior cloning (BC) in robotic manipulation has been largely confined to tasks where demonstrations can be effectively collected through human teleoperation. However, demonstrations for contact-rich manipulation tasks that require complex coordination of multiple contacts are difficult to collect due to the limitations of current teleoperation interfaces. We investigate how to leverage model-based planning and optimization to generate training data for contact-rich dexterous manipulation tasks. Our analysis reveals that popular sampling-based planners like rapidly exploring random tree (RRT), while efficient for motion planning, produce demonstrations with unfavorably high entropy. This motivates modifications to our data generation pipeline that prioritizes demonstration consistency while maintaining solution diversity. Combined with a diffusion-based goal-conditioned BC approach, our method enables effective policy learning and zero-shot transfer to hardware for two challenging contact-rich manipulation tasks.

Open Source Robotics