README.md

Differentiable robot-object interaction for learning object properties

This repository is the official implementation of the paper:

Learning Object Properties Using Robot Proprioception via Differentiable Robot-Object Interaction
Peter Yichen Chen, Chao Liu, Pingchuan Ma, John Eastman, Daniela Rus, Dylan Randle, Yuri Ivanov, Wojciech Matusik
MIT CSAIL, Amazon Robotics, University of British Columbia
International Conference on Robotics and Automation (ICRA), 2025

A big shoutout to the Nvidia Warp team! Warp integrates effortlessly with Torch, streamlining the use of differentiable simulation for Torch-based optimization workflows.

Installation

Install the required packages first:

pip install -r requirements.txt

For visualization, install these optional packages:

• bpy
• blendertoolbox

and these softwares:

• Blender
• ffmpeg

Usage

To calibrate object properties, use the following command:

python train.py --config-name hard_ball

To evaluate the calibrated object property, use the following command:``

python eval.py --config-name hard_ball ckpt=experiments/log/robotis_2_hard_ball/open_manipulator/open_manipulator_joint2_only_v2/train/training_stats.pt ckpt_idx=8

To visualize the robot, use the following command:

python render_usd.py --usd-path experiments/log/robotis_2_hard_ball/open_manipulator/open_manipulator_joint2_only_v2/test/test_ckpt_idx_0008.usd

If this helps you, please consider citing the paper below.

@misc{chen2025learningobjectpropertiesusing,
      title={Learning Object Properties Using Robot Proprioception via Differentiable Robot-Object Interaction}, 
      author={Peter Yichen Chen and Chao Liu and Pingchuan Ma and John Eastman and Daniela Rus and Dylan Randle and Yuri Ivanov and Wojciech Matusik},
      year={2025},
      eprint={2410.03920},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2410.03920}, 
}