Maryland model learns motion traces from video to guide robot tasks

What happened

In physical tests with a robot arm and two-finger gripper, the system averaged 91.7% success across placing objects in a sink, pouring almonds and unfolding a towel. Simulation results across eight kitchen tasks were mixed: its 30.25% average exceeded one comparison system but trailed another.

Why it matters

The approach separates learning what should move from learning how a specific robot should execute the motion, which may reduce reliance on hardware-specific data. It is not a general robot policy or deployment: the action expert still needs robot demonstrations, the model does not explicitly represent force or touch, and broader robot and task testing remains future work.

Source checked October 31, 2026. Status describes the evidence in the linked source at the time of review.