Batched Differentiable Rigid Body Dynamics in PyTorch for GPU-Accelerated Robot Learning

Researchers introduced BARD, a self-contained PyTorch engine for articulated rigid-body dynamics tailored to batched GPU evaluation and automatic differentiation.

By using caching, precomputed joint transforms, and level-parallel propagation, BARD significantly improves throughput while preserving numerical accuracy.

Benchmarks on multiple robot models show large speedups over Pinocchio, especially on NVIDIA H200 hardware.

The system is also demonstrated in gradient-based system identification and an Isaac Lab training pipeline, helping remove a key bottleneck in robot learning.