MIT CSAIL's UAN bridges sim-to-real to train robots that throw and lift

EpisodeYang · x · 2026-10-20

MIT CSAIL researchers (Nolan Fey, Pulkit Agrawal, et al.) released UAN (Unsupervised Actuator Net) for athletic loco-manipulation.

• Replaces traditional trajectory-tracking rewards with task rewards ("throw the ball as far as you can", "lift the weight as fast as possible"), tackling reward hacking and under-directed exploration.
• UAN uses real-world data to correct actuator errors in simulation (harmonic drive vs. QDD motor friction and other unmodeled dynamics) without torque sensing, keeping behaviors robust and transferable.
• A pre-training/fine-tuning pipeline resets the simulator to recorded hardware states, replays recorded commands, and learns command corrections to match real trajectories; reference trajectories guide exploration.
• The robot lifts, throws, and drags with high sim-to-real fidelity; a viral clip shows the arm flying off with the ball, attributed to massive closed-loop training.

Original post →

More from Embodied

• ScaleNav pretrains robot navigation on internet egocentric videos — Yihui Yang · 2026-10-20
• Oura ring misclassifies Indian folk dance Dandia/Garba as cross-country skiing — salgar · 2026-10-20
• whurley logs nearly 100 Tesla CyberCab robotaxi rides — whurley · 2026-10-20
• Robotics take: IK guarantees no precision edge, hardware—not the solver—is the bottleneck — ChongZzZhang · 2026-10-20
• The most important AI device of the next decade won't be a smartphone — VraserX · 2026-10-20
• NUS MAGIC Lab is hiring a postdoc, 2 PhD students and 3 research assistants for robotics — DJiafei · 2026-10-20

Embodied channel →