Jim Fan (NVIDIA) warns robot hardware reliability slows software iteration, calls for new benchmarks and pretraining shift

Jim Fan (NVIDIA) summarizes three 2025 lessons: (1) modern humanoid hardware often outperforms current AI control but fragile hardware (overheating, motor/firmware failures) severely limits software iteration; (2) robot benchmarks are fragmented, non-reproducible and prone to cherry-picked demos; (3) VLM-based VLA approaches mismatch robotic needs—he advocates video-world-model pretraining to learn dynamics and fine visual details instead. He urges the community to standardize benchmarks by 2026 and rethink pretraining objectives.