On the Numerical Reliability of Differentiable Physics-Based Optimization for Robotic Material Manipulation

The paper investigates the numerical reliability of gradients in differentiable physics-based optimization for robotic material manipulation. Using two Material Point Method benchmarks, it identifies three key issues: GPU many-to-one sums that alter gradient signs, reduced reliability of finite-difference checks for long rollouts, and the impact of observation and loss definitions on optimization outcomes. The study recommends reproducible accumulation, finite-difference validation, and explicit objective reporting to improve robustness in robotic optimization.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at Hugging Face Trending Papers.

• robotics
• benchmarks