Contact Points Discovery for Soft-Body Manipulations with Differentiable Physics

Authors: Sizhe Li, Zhiao Huang, Tao Du, Hao Su, Joshua B. Tenenbaum, Chuang Gan

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results suggest that: 1) on multi-stage tasks that are infeasible for the vanilla differentiable physics solver, our approach discovers contact points that efficiently guide the solver to completion; 2) on tasks where the vanilla solver performs sub-optimally or near-optimally, our contact point discovery method performs better than or on par with the manipulation performance obtained with handcrafted contact points.
Researcher Affiliation Collaboration Sizhe Li University of Rochester sli96@u.rochester.edu; Zhiao Huang UC San Diego z2huang@eng.ucsd.edu; Tao Du MIT taodu@csail.mit.edu; Hao Su UC San Diego haosu@eng.ucsd.edu; Joshua B. Tenenbaum MIT BCS, CBMM, CSAIL jbt@mit.edu; Chuang Gan MIT-IBM Watson AI Lab ganchuang@csail.mit.edu
Pseudocode Yes C PSEUDOCODE OF CPDEFORM; Algorithm 1 CPDeform; D CONTROLLER OPTIMIZATION WITH DIFFERENTIABLE PHYSICS; Algorithm 2 Differentiable Physics Solver for Controller Optimization
Open Source Code No Demos are available on our project page1. 1Project Page: http://cpdeform.csail.mit.edu. This is a project page and does not explicitly state that the source code for the methodology is provided there, nor is it a specific code repository.
Open Datasets Yes To extensively evaluate our method, we introduce Plasticine Lab-M, a dataset that extends the existing differentiable physics benchmark Plasticine Lab with seven new challenging multi-stage soft-body manipulation tasks, and contains the multi-stage environment Pinch in Plasticine Lab.
Dataset Splits No The paper does not provide explicit details about training, validation, and test dataset splits (e.g., percentages, sample counts, or specific files). It mentions episodes and steps for optimization/training but not data partitioning.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU models, CPU types, or cloud resources with specs) used to run its experiments.
Software Dependencies No The paper mentions software like 'differentiable physics solver' and 'SAC', 'PPO', 'TD3' but does not provide specific version numbers for any software components or libraries.
Experiment Setup Yes For each stage, we optimize for 200 episodes for differentiable physics-based approaches with a learning rate 0.1. For each environment, we modestly choose a horizon of 10 or 20. We restrict the number of environment steps used for optimization under 1 million. We train each RL algorithm on each environment for 1000 episodes, with 1000 environment steps per episode, which accounts for the 1 million environment-step limit.