[Paper Review] Contact Points Discovery for Soft-Body Manipulations with Differentiable Physics

Sizhe Li, Zhiao Huang+4 more|arXiv (Cornell University)|May 5, 2022

Robot Manipulation and Learning4 citations

TL;DR

This paper proposes CPDeform, a differentiable physics framework that uses optimal transport-based contact point discovery to overcome local minima in soft-body manipulation. By iteratively identifying high-priority contact points for end-effectors, it enables effective single- and multi-stage deformation tasks, significantly improving success rates on challenging benchmarks where vanilla differentiable physics fails.

ABSTRACT

Differentiable physics has recently been shown as a powerful tool for solving soft-body manipulation tasks. However, the differentiable physics solver often gets stuck when the initial contact points of the end effectors are sub-optimal or when performing multi-stage tasks that require contact point switching, which often leads to local minima. To address this challenge, we propose a contact point discovery approach (CPDeform) that guides the stand-alone differentiable physics solver to deform various soft-body plasticines. The key idea of our approach is to integrate optimal transport-based contact points discovery into the differentiable physics solver to overcome the local minima from initial contact points or contact switching. On single-stage tasks, our method can automatically find suitable initial contact points based on transport priorities. On complex multi-stage tasks, we can iteratively switch the contact points of end-effectors based on transport priorities. To evaluate the effectiveness of our method, we introduce PlasticineLab-M that extends the existing differentiable physics benchmark PlasticineLab to seven new challenging multi-stage soft-body manipulation tasks. 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.

Motivation & Objective

• To address the challenge of local minima in differentiable physics solvers caused by suboptimal or fixed initial contact points during soft-body manipulation.
• To enable effective multi-stage soft-body manipulation requiring contact point switching, which standard solvers often fail to complete.
• To develop a contact point discovery method that is principled, geometry-based, and compatible with existing differentiable physics pipelines.
• To evaluate the method on a new benchmark, PlasticineLab-M, to demonstrate performance gains over handcrafted and random contact points.
• To show that contact point discovery can be used as a standalone heuristic to improve solver performance on single-stage tasks.

Proposed method

• Integrates optimal transport (Wasserstein distance) to compute transport priorities between current and target shapes, identifying high-impact contact points for deformation.
• Uses these transport priorities to guide end-effector placement, effectively simulating human-like iterative contact-and-deform behavior.
• Applies a heuristic iterative switching mechanism for multi-stage tasks, allowing the solver to escape local minima by repositioning contact points based on shape progress.
• Combines the contact discovery module with a stand-alone differentiable physics solver to enable gradient-based optimization on soft-body deformation tasks.
• Employs a particle-based representation of soft bodies to compute shape differences and transport costs efficiently.
• Leverages the PlasticineLab differentiable physics simulator for trajectory optimization and gradient computation.

Experimental results

Research questions

• RQ1 Can optimal transport-based contact point discovery improve the success rate of differentiable physics solvers in soft-body manipulation tasks?
• RQ2 Can the proposed method automatically identify suitable initial contact points for single-stage tasks without human intervention?
• RQ3 Can the framework handle complex multi-stage tasks requiring contact point switching, where vanilla solvers fail due to local minima?
• RQ4 How does the performance of the contact discovery method compare to handcrafted or randomly selected contact points on single-stage tasks?
• RQ5 To what extent can iterative contact point switching via transport priorities enable completion of long-horizon, multi-stage soft-body manipulation tasks?

Key findings

• On multi-stage tasks that are infeasible for the vanilla differentiable physics solver, CPDeform successfully completes all seven new tasks introduced in PlasticineLab-M, demonstrating robustness to contact switching.
• On single-stage tasks where the vanilla solver performs sub-optimally or near-optimally, the contact point discovery method alone outperforms or matches the performance of handcrafted contact points.
• The optimal transport-based contact discovery method achieves performance comparable to or better than human-defined contact points without requiring task-specific tuning.
• The iterative contact-and-deform strategy enables the solver to escape local minima by dynamically repositioning contact points based on shape progress.
• The method generalizes across different soft-body shapes and does not require material-specific parameters, relying only on shape and topology information.
• The framework shows that geometric analysis via optimal transport can serve as an effective heuristic for contact point selection in differentiable physics.