SO-101 Bench: a $120 real2sim2real robot benchmark where sim rankings transfer to real hardware

chris_j_paxton · x · 2026-10-29

SO-101 Bench v1 is out—an accessible ($120 robot), reproducible real2sim2real robot learning benchmark.

Key finding: policies trained only in simulation keep their rankings on the real robot, meaning the simulation correlates with real-world performance and can serve as a reliable evaluation proxy.
On the primary eval, an agent + VLA combo beat every fine-tuned VLA/WAM.
Chris Paxton calls the combination of reproducible hardware benchmark, associated simulation, and sim-real correlation 'really useful'.

More from Embodied