In June 2019, a team from MIT’s Computer Science and Artificial Intelligence Laboratory published ChainQueen, a simulator built specifically for soft robots — machines made from flexible, deformable materials rather than rigid components. The MIT CSAIL post describes the system and its potential applications.
The problem with simulating soft robots
Physics simulators are standard tools in robotics: they model how a robot’s actions will affect the physical world before those actions are carried out. Existing simulators work reasonably well for rigid robots. For soft robots — those made from elastic, deformable materials — the underlying physical laws are substantially more complex, requiring considerably more computational power to simulate accurately.
The MIT CSAIL post describes ChainQueen as a new simulator designed specifically for soft robots, demonstrated across an eclectic range of robotic designs including a crawling robot and a four-legged running robot.
Technical approach
ChainQueen builds on the Material Point Method, a technique for simulating material properties that the post notes was previously used to create realistic snow in the film “Frozen.” Specifically, the simulator uses a faster variant called MLS-MPM — Moving Least Squares Material Point Method — developed by paper co-author Yuanming Hu in collaboration with researchers at the University of Pennsylvania and UCLA. MLS-MPM can be more readily incorporated into inference, control, and co-design systems than earlier MPM variants.
The key innovation is that ChainQueen makes MLS-MPM differentiable — meaning the derivative, or direction of steepest descent, can be computed for all aspects of robot control and design. This allows numerical optimizers to efficiently search for optimal robot configurations, a task the post states can be done “much faster than recent derivative-free approaches, such as reinforcement learning.”
The name “ChainQueen” refers to the chain rule of calculus, which underlies the system’s feedback computation. The team developed a high-performance GPU implementation.
Design feedback
Beyond evaluating robot designs, ChainQueen provides feedback on how designs can be improved. The post describes this as a distinction from existing simulation tools that could evaluate but not guide design iteration.
Andrew Spielberg, a PhD student and co-author, is quoted in the post: “We believe this system has the potential to dramatically accelerate the development of soft robots. We’ve also created a TensorFlow interface that will allow users at all levels to develop their own soft robotics systems without needing to understand the simulator’s low-level details.”
The paper was presented at the IEEE International Conference on Robotics and Automation. Co-authors include Hu, Spielberg, visiting student Jiancheng Liu, PhD student Jiajun Wu, and MIT professors Joshua B. Tenenbaum, William T. Freeman, Daniela Rus, and Wojciech Matusik.
Scope and future work
The post notes the project focuses on elastic designs. Hu said future work could extend the simulator to other materials — plastics, cloth, fluids — and to more complex interactions with rigid and soft environments. The team indicated a plan to make the GPU implementation open-source.