DELUCA-Differentiable Control Library:Environments, Methods and Benchmarking

Reviewer 1: Very relevant. The paper presents on open-source library of differentiable physics and robotics environments, and also provides gradient-based control methods and a benchmarking suite.

I think that’s a very important direction and would help control and physics engine design communities to come together to offer a wide range of tools to enable fast gradient based learning of control policies. This should help various communities in control, sim-to-real as well as simulation and design. “

Reviewer 2: Very relevant. The paper presents an open-source library of differentiable physics (and robotics). The library also implements gradient-based control methods, and a benchmarking suite that links it to control suites like OpenAI Gym. To this end, the library uses auto-differentiation methods provided by Jax. The paper demonstrates usage of the new library on a ventilator control task and on an inverted pendulum task.

Reviewer 3: Highly relevant to the themes of this workshop. Paper proposes to build a differentiable system to accommodate time varying linear dynamical system and integrate with various RL benchmarking suites for control tasks.

The shortcoming of current paradigms are well motivated with respect to policy gradients and where are how having differentiable dynamics tied with the rest of simulation will add value.

Preliminary experiments on pendulum and planar quadrotor show promising results and timing improvements are excellent. Would like to see a discussion on how far this can be pushed and what are the current technical hurdles. For instance, for high dof robot manipulation or locomotion problems. Is it just a matter of implementation and solid engineering?

A list or table would help in summarizing what is supported, in development, and planned. Would probably help the community from reinventing the wheel and use this library (once released).