Abstract
Differentiable physics has recently been shown as a powerful tool for solvingsoft-body manipulation tasks. However, the differentiable physics solver oftengets stuck when the initial contact points of the end effectors are sub-optimalor when performing multi-stage tasks that require contact point switching,which often leads to local minima. To address this challenge, we propose acontact point discovery approach (CPDeform) that guides the stand-alonedifferentiable physics solver to deform various soft-body plasticines. The keyidea of our approach is to integrate optimal transport-based contact pointsdiscovery into the differentiable physics solver to overcome the local minimafrom initial contact points or contact switching. On single-stage tasks, ourmethod can automatically find suitable initial contact points based ontransport priorities. On complex multi-stage tasks, we can iteratively switchthe contact points of end-effectors based on transport priorities. To evaluatethe effectiveness of our method, we introduce PlasticineLab-M that extends theexisting differentiable physics benchmark PlasticineLab to seven newchallenging multi-stage soft-body manipulation tasks. Extensive experimentalresults suggest that: 1) on multi-stage tasks that are infeasible for thevanilla differentiable physics solver, our approach discovers contact pointsthat efficiently guide the solver to completion; 2) on tasks where the vanillasolver performs sub-optimally or near-optimally, our contact point discoverymethod performs better than or on par with the manipulation performanceobtained with handcrafted contact points.