Can a robot's value function design better robot bodies?

Instead of running expensive reinforcement learning loops for each new robot design, researchers train a single embodiment-aware value function across many robot morphologies, then reuse it as a differentiable optimizer for new body shapes. Testing on up to 50 robot designs and design spaces exceeding 1100 parameters, the approach both optimizes complete embodiments and identifies which physical parameters limit performance—cutting the iteration cost for robot design.