Differentiable Co-Optimization of Mass-Limited Legged Robots Across Variable Gravity

Anais Jeger, Leone Costi, Giacomo Acciarini, Dario Izzo

Keywords: Space Robotics, Differentiable, Co-Design

TL;DR: A differentiable co-design framework jointly optimizes robot morphology and gait across different gravity levels, revealing gravity-specific design trade-offs for efficient small legged robots on planetary bodies under a fixed mass budget.

Abstract:

We utilize a differentiable co-design framework to study legged locomotion across
widely varying gravitational environments, with the objective of gaining insights
into the design of small, mass-limited robots able to operate efficiently on
planetary bodies and moons. Our method is based on end-to-end gradient-based
optimization through rigid-body contact dynamics, leveraging NVIDIA Warp's
GPU-accelerated automatic differentiation to tune morphology and gait parameters
jointly over long-horizon simulations under a fixed 12kg mass budget. Our
results demonstrate that treating gravity as a first-class design variable and
co-optimizing morphology with gait coordination produces locomotion solutions
with gravity-dependent trade-offs in body-leg mass allocation and limb geometry,
providing a practical tool for destination-specific legged robot design.

Submission Number: 15