阅读笔记：PlasticineLab A Soft-Body Manipulation Benchmark with Differentiable Physics

academical expressions

• fall short on … (eg. malti-stage task)
• physics-based skill learning task
• is highlighted by the adoption of …
• make it a representative medium to …
• An incomplete list includes …
• While some support …, none has provided …
• We fill in this gap with our …
• Closely related to our work is …
• is originated from … but with significant modifications in order to add our novel support for …
• an extensively studied topic for years
• by perturbing …
• We refer the reader to … for more details on …
• We evalluate the algorithms on xxx configurations and report an algebraic average score.
• revealing their deficiency in …
• We speculate that our environment favors xxx over algorithms dependent on …

Introduction

It’s a pity that there are many useful engines but not for soft-body because of complexity.
We overcome difficulty and add new plasticine material and new differentiable feature.
How we implement and model and what exams we did.

Contributions:

①1st skill learning benchmark and both sof materials;
②fully-featured diff- engine;
③ broad tasks enable systematic eval & anlys of RL & grad-based planning algs.

Related Work

Learning in virtual environments

many platforms, but no assets or tasks for soft-body manipulation

Differentiable physics engine

approximate physical simulators using NN
physics-based simulators using differentiable programming systems
others (???) LCP, QR decompositions
ChianQueen & DiffTaichi √

Trajectory optimization

High DoF and intricate contacts make motion planning for soft-body manipulation under exploration in trajectory optimization and diff-physics, but we use the new tech in DiffTaichi, unlocking grediant-based optim, to solve.

Learning-based soft body manipulation

different 1: propose a comprehensive benchmark for evaluating and developing novel algorithms in soft-body research
different 2: study soft body manipulation with rigid manipulators

The PlasticineLab Learning Environment

Task representation

MDP:

• state: $N_p\times d_p$ (position + velocity + deformation gradient + affine velocity field, flatten), $N_m\times d_m$ (position + orientation if any)
• observation: $N_k\times 6$ (position + velocity) downsample fixed relative location particles if initial configuration for the same task, $N_m\times d_m$