Title: Shaping Robotic Forms: Leveraging Value Gradients for Multi-Embodiment Design

Abstract: This study introduces a methodology that transforms generalist, multi-embodiment value functions into versatile tools for robotic design. Rather than initiating a distinct reinforcement learning co-design cycle for every individual robot, we first develop an embodiment-aware policy and value function encompassing a wide array of robot configurations. Once training is complete, the resulting value function is frozen and employed as a differentiable surrogate, allowing for the optimization of potential embodiments via value gradients. We validate this approach across various design scenarios, ranging from modified single robots to unseen models across different morphological categories. Our models were trained on up to 50 robots, navigating design spaces defined by more than 1,100 continuous embodiment parameters. In addition to refining full embodiments, we demonstrate that value gradients can pinpoint specific design and control parameters that constrain performance, thereby facilitating both the optimization and the analytical assessment of novel robotic designs.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC