Abstract
Humanoid robots could benefit from using their upper bodies for support contacts, enhancing their workspace, stability, and ability to perform contact-rich and pushing tasks. In this paper, we propose a unified approach that combines an optimization-based multi-contact whole-body controller with Flow Matching, a recently introduced method capable of generating multi-modal trajectory distributions for imitation learning. In simulation, we show that Flow Matching is more appropriate for robotics than Diffusion and traditional behavior cloning. On a real full-size humanoid robot (Talos), we demonstrate that our approach can learn a whole-body non-prehensile box-pushing task and that the robot can close dishwasher drawers by adding contacts with its free hand when needed for balance. We also introduce a shared autonomy mode for assisted teleoperation, providing automatic contact placement for tasks not covered in the demonstrations.
Multi-Support Experiments on Talos Humanoid Robot
Non-Prehensile Box Pushing (Autonomous)
Closing of Dishwasher Drawers (Autonomous)
When trained on the red T-box, the autonomous policy performs poorly out-of-distribution with the blue U-box
This is mitigated with Shared Autonomy / Assisted Teleoperation approach
The human operator commands the left hand, and the policy commands the right hand
Assisted Teleoperation Pushing the Blue U-Box
Assisted Teleoperation Closing Dishwasher Drawers
Methodology
Multi-Support Manipulation Tasks:
- Use additional contact to extend manipulation capabilities
- Reach further, be more stable, apply higher pushing forces
- Leverage whole-body motion and multi-contact strategies
Main Ideas
- Imitation learning from human demonstrations (behavior cloning)
- Use human "common sense" to learn contact sequence and location
- Solve out-of-distribution tasks not covered by demonstrations using shared autonomy assisted teleoperation
Imitation Learning with Flow Matching
- The policy outputs a full command trajectory instead of a single action
- Use Flow Matching, a generative method able to capture high-dimensional and multi-modal distributions
- Simpler framework, slightly more performant than Diffusion for robotics grounded in optimal transport theory
Multi-Contact Low-Level Controller
-
Use our previous work SEIKO Retargeting and Controller
https://hucebot.github.io/seiko_controller_website/ - Retargeting from Cartesian to whole-body commands with feasibility constraints
- Implement smooth contact switch
- Regulate contact forces on our position-controlled robot
Comparison Between Flow Matching and Diffusion Methods
Simulated T-Box Pushing Task
Simulated Bimodal Contact Reaching Task
BibTeX
@inproceedings{rouxel2024flowmultisupport,
title={Flow Matching Imitation Learning for Multi-Support Manipulation},
author={Rouxel, Quentin and Ferrari, Andrea and Ivaldi, Serena and Mouret, Jean-Baptiste},
booktitle={2024 IEEE-RAS 23rd International Conference on Humanoid Robots (Humanoids)},
year={2024},
organization={IEEE}
}