FADA (CoRL 2026) open-sourced: humanoid robots adapt to new conditions from 2 minutes of experience
GuanyaShi · x · 2026-09-12
- LeCAR-Lab's FADA, a few-shot domain adaptation framework for humanoid control, is accepted to CoRL 2026 and now open-sourced on GitHub.
- Adapts a trained policy to new real-world dynamics (heavy payload, soft floor, steep slope, unfamiliar terrain) using only 2 minutes of target-domain data.
- No rewards, no demonstrations, no motion capture, no simulator retuning — it keeps the task goal and changes only how the robot moves.
- Repo includes the full pipeline: oracle PPO training + Planner-IDM DAgger distillation, IDM LoRA finetuning, MuJoCo sim2sim deployment/data collection, and sim2real real-robot deployment.
More from Embodied
- Edinburgh Spinout Singular Photonics Launches Litavis, a SPAD Sensor That Processes Photons In-Pixel — Scobleizer · 2026-09-12
- Open-source robot arms now sell for $999 with 3.3 lb payload and 0.2mm accuracy — Scobleizer · 2026-09-12
- Tesla Europe recaps a year of FSD Supervised testing across Swedish streets and Nordic weather — yunta_tsai · 2026-09-12
- Ex-OpenAI RLHF researcher's robotics startup ships open nav model LightNav-0 and resilient control LightREACT in 10 days — 机器之心 · 2026-09-12
- whurley jokes that Tesla's robotaxi account came in a little strong — whurley · 2026-09-12
- Rumor: OpenAI quietly expanding its humanoid robot project — DJiafei · 2026-09-12