PlasticineLab: A benchmark for manipulating soft bodies with differentiable physics

Simulated virtual environments help develop and evaluate intelligent agent algorithms. Most of the existing ones, however, focus on rigid body dynamics, although soft body environments can be widely applied. For example, soft bodies can be used to simulate virtual operations or to develop biomimetic actuators in robotics. A recent article therefore proposes a framework for performing and evaluating ten soft body manipulation tasks.

The tasks require complex operations including rolling, chopping or shaping. For example, in one of the tasks the agent has to wind a rope, modeled as a long piece of plasticine, around a rigid column using two spherical manipulators. The framework uses differentiable physics and provides analytical gradient information that can be used in supervised learning with gradient-based optimization. The study enables reinforcement learning and gradient-based planning algorithms to be compared.

Simulated virtual environments are one of the main driving forces behind the development and evaluation of skills learning algorithms. However, existing environments usually only simulate rigid body physics. In addition, the simulation process usually does not provide any color gradients that could be helpful for planning and controlling optimizations. We're introducing a new differentiable physics benchmark called PasticineLab, which includes a diverse collection of soft body manipulation tasks. For each task, the agent uses manipulators to deform the plasticine into the desired configuration. The underlying physics engine supports the differentiable elastic and plastic deformation with the help of the DiffTaichi system and presents robot agents with many as yet unexplored challenges. We use this benchmark to evaluate several existing RL methods (reinforcement learning) and gradient-based methods. Experimental results suggest that 1) RL-based approaches have difficulty solving most problems efficiently; 2) Gradient-based approaches can quickly find a solution within ten iterations by optimizing control sequences with the integrated differentiable physics engine, but still cannot fall back on multi-level tasks that require long-term planning. We expect that PlasticineLab will encourage the development of novel algorithms that combine differentiable physics and RL for more complex physics-based skills learning tasks.

Link to research article: Huang, Z. et al., "PlasticineLab: A soft-body manipulation benchmark with differentiable physics"2021.

Link to the project page:

<! –

comment these news or articles


Leave a Reply

Your email address will not be published. Required fields are marked *