Why robot learning stays slow: speeding up physics on GPUs

Robot learning with reinforcement learning requires computing physics millions of times during training, but popular libraries like Pinocchio run on CPUs and bottleneck GPU pipelines. BARD implements rigid-body dynamics directly in PyTorch with three tricks—lazy evaluation, precomputed transforms, and parallel tree propagation—to run on GPUs with full automatic differentiation. On realistic 7–23 degree-of-freedom robots, it matches Pinocchio's accuracy while reaching 64× throughput gain for kinematics and 8.5× speedup in training 4000+ parallel environments.