Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Huaijiang Zhu, Tong Zhao, Xinpei Ni, Jiuguang Wang, Kuan Fang, Ludovic Righetti, Tao Pang

TL;DR

The paper tackles the challenge of learning contact-rich dexterous manipulation from demonstrations, which are hard to obtain via teleoperation. It shows that planning-based data, especially from traditional sampling planners like RRT, can be high-entropy and harmful for BC, and proposes a data-generation pipeline that emphasizes demonstration consistency using greedy and PRM-based planners. By pairing this planning-driven data with a diffusion-based goal-conditioned BC, the authors demonstrate effective policies and zero-shot hardware transfer on AllegroHand and IiwaBimanual tasks, highlighting the practical potential of model-based planning to scale BC beyond simple grasping. The work also documents limitations, including sim-to-real gaps and challenges with non-rigid objects, pointing to future avenues like hybrid sim-real training and corrective data strategies to further close the sim-to-real gap.

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.

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

TL;DR

The paper tackles the challenge of learning contact-rich dexterous manipulation from demonstrations, which are hard to obtain via teleoperation. It shows that planning-based data, especially from traditional sampling planners like RRT, can be high-entropy and harmful for BC, and proposes a data-generation pipeline that emphasizes demonstration consistency using greedy and PRM-based planners. By pairing this planning-driven data with a diffusion-based goal-conditioned BC, the authors demonstrate effective policies and zero-shot hardware transfer on AllegroHand and IiwaBimanual tasks, highlighting the practical potential of model-based planning to scale BC beyond simple grasping. The work also documents limitations, including sim-to-real gaps and challenges with non-rigid objects, pointing to future avenues like hybrid sim-real training and corrective data strategies to further close the sim-to-real gap.

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.

Paper Structure

This paper contains 27 sections, 1 equation, 4 figures, 2 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Method
  3. Planning-Guided Data Synthesis
  4. Planning through contact
  5. Simulation Rollout
  6. Goal-Conditioned Behavior Cloning
  7. Data Curation and Planner Modifications
  8. Bimanual Manipulation
  9. Random initialization
  10. Training parameters
  11. Success criteria
  12. Planner design
  13. Performance analysis
  14. In-Hand Re-Orientation
  15. Planner modifications
  16. ...and 12 more sections

Figures (4)

  • Figure 1: Framework overview.
  • Figure 2: Entropy of linear and angular velocity directions of the RRT and greedy datasets for (a) IiwaBimaual (IB) and (b) AllegroHand (AH). White indicates that there is no data.
  • Figure 3: Example demonstrations for IiwaBimanual (IB) and AllegroHand(AH). In all subfigures, the solid frames indicate the goal object configuration. For both tasks, contact RRT covers more space by following a more meandering path before reaching the goal than their lower-entropy counterparts.
  • Figure 4: Normalized weighted distance to goal of the object is plotted against trajectory completion percentage for (a) contact RRT IiwaBimanual (IB) and (b) greedy search IB. Each colored curve represents one demonstration trajectory. Dots along a curve represent regrasps. Gray bars represent regrasp entropy for discretized time intervals. (c) Histograms of distance progress per contact segment for IB. (d)-(f) Similar plots for the AllegroHand (AH) task. We show 100 trajectories for each task to avoid cluttering.