Abstract
The current dominant paradigm for robotic manipulation involves two separatestages: manipulator design and control. Because the robot's morphology and howit can be controlled are intimately linked, joint optimization of design andcontrol can significantly improve performance. Existing methods forco-optimization are limited and fail to explore a rich space of designs. Theprimary reason is the trade-off between the complexity of designs that isnecessary for contact-rich tasks against the practical constraints ofmanufacturing, optimization, contact handling, etc. We overcome several ofthese challenges by building an end-to-end differentiable framework forcontact-aware robot design. The two key components of this framework are: anovel deformation-based parameterization that allows for the design ofarticulated rigid robots with arbitrary, complex geometry, and a differentiablerigid body simulator that can handle contact-rich scenarios and computesanalytical gradients for a full spectrum of kinematic and dynamic parameters.On multiple manipulation tasks, our framework outperforms existing methods thateither only optimize for control or for design using alternate representationsor co-optimize using gradient-free methods.