Contact-UAN: MIT's residual world model fixes humanoid sim-to-real with 6.5 minutes of data

kevin_zakka · x · 2026-10-17

• MIT CSAIL researchers (Nolan Fey, Khai Nguyen, et al., Pulkit Agrawal's lab) introduce Contact-UAN, a residual world model for humanoid control, accepted at CoRL 2026 with a Best Paper Award at the RSS 2026 Robot World Models workshop.
• The problem: humanoid policies transfer qualitatively to hardware but lose force interactions due to actuator sim-to-real gaps. Contact-UAN preserves the simulator's physics priors while correcting the actuator command channel using only onboard proprioception, with contact-consistent replay preventing the model from learning reset artifacts.
• Trained on just 6.5 minutes of real walking data, it generalizes to force/impulse-sensitive tasks: 43% lower force-tracking error and 20% more torque efficiency — no torque sensing or motion capture needed. Less domain randomization, higher jumps, stable landings.

Original post →

More from Embodied

• OpenAI engineer used his Dot to triage a 3am on-call incident via a Codex thread from his phone — OpenAIDevs · 2026-10-17
• SoftBank reportedly weighs buying Allied Universal at up to $18B for AI robots — SuB8u · 2026-10-17
• Single-video 3D reconstruction via agent→MCP→Blender simulates industrial hazards — RexDouglass · 2026-10-17
• ROS 2 LiDAR localization runs 390 m outdoors on handheld Livox MID-360 — rsasaki0109 · 2026-10-17
• Embodied Astra tops RoboMME at 92.1% — beating humans by just keeping full visual history in context — hankyang94 · 2026-10-17
• 100% blind developer built AccessSlice with Claude to make 3D printing accessible — Mrblindguardian · 2026-10-17

Embodied channel →