Designing a robot arm's morphology—its link lengths and joint placement—often relies on Yoshikawa's Manipulability Index, a classical measure of kinematic dexterity. But this index is typically evaluated in simulation as a purely geometric quantity, ignoring the real-world effort required to move and hold the arm.
This paper instead frames morphology optimization as a reinforcement learning problem. The agent is rewarded for kinematic performance but also penalized based on measured effort from physical hardware, bridging the gap between simulated design and actual operation.
The authors validate their approach on a planar 2R manipulator, showing that hardware-validated RL can yield morphologies that differ from those chosen by pure kinematic optimization. The work highlights the importance of incorporating physical effort early in the design loop, rather than treating it as an afterthought.