A2-MEAS framework lifts household robot task success with learned correction codes

Researchers led by Chao Liang have described A2-MEAS in an arXiv paper, a multi-agent framework for robots in which learned correction codes mix corrective language, spatial evidence, the history of what the robot did and proposed actions. A component called MetaBrain supervises training and guides the robot at run time, and the codes can change its behaviour at deployment without retraining.

On all 100 public tasks of the BEHAVIOR 2026 benchmark, A2-MEAS reached a Q-score of 0.6411 and 45.40% task success. Its action model, A2-WLA, averaged 93.98% success on RoboTwin 2.0 and 57.92% across 12 real-robot tasks in five household settings, against 26.67% for pi0.5. The team plans to release the system, models and training tools.