Toward the Era of Agentic Embodied Intelligence
A robot correction system raised completion after failures at two coffee-filter steps from 14% to 94% and from 8% to 82%.
The authors’ A2-MEAS framework combines corrective language, spatial evidence and execution history to guide robot actions, including deployment-time adjustments with frozen model weights. The coffee results compare its history-conditioned adapter with retries by the underlying A2-WLA model. Across 12 real-robot tasks in five household settings, A2-WLA reported 57.92% average success, versus 26.67% for pi0.5. The authors plan to release the system, models and training infrastructure.
HF Daily Papers' note
The authors’ A2-MEAS framework combines corrective language, spatial evidence and execution history to guide robot actions, including deployment-time adjustments with frozen model weights. The coffee results compare its history-conditioned adapter with retries by the underlying A2-WLA model. Across 12 real-robot tasks in five household settings, A2-WLA reported 57.92% average success, versus 26.67% for pi0.5. The authors plan to release the system, models and training infrastructure.
HF Daily Papers' note
score 5