Paper briefing for humanoid robotics readers.

NVIDIA's COMPASS uses residual reinforcement learning and policy distillation to deploy a single navigation policy across wheeled and legged platforms — with zero-shot sim-to-real transfer on the Unitree G1 humanoid.

HEX is a 2.4B-parameter open-source VLA model pretrained on 12 million frames from seven different humanoid robots — and it controls all of them with a single policy.

A survey of the most significant papers on bipedal locomotion from 2025 — reinforcement learning, model predictive control, and sim-to-real transfer.