GPUSimBench reveals hidden limits of GPU-based robot simulators

A new benchmark called GPUSimBench, accepted at IROS 2026, systematically exposes hidden limitations in mainstream GPU-accelerated robotic simulators like NVIDIA's Isaac Lab and Genesis. Developed by Huzhenyu Zhang and colleagues, the work focuses on three critical axes: scalability, physical consistency, and computational determinism. The benchmark uses a controlled inclined-plane task to measure how closely simulated dynamics match real-world physics, then scales environment counts to measure throughput and memory footprints.

The most striking finding is the inherent non-determinism introduced by GPU-batched execution. Even under identical initial conditions, the same simulator can produce different results across runs and across parallel environments. The authors identify four empirical regimes of stochasticity, warning that unbounded scaling can compromise reproducibility without explicit constraints. This has direct implications for training reliable robot learning models at scale.