A2-MEAS embodied framework reports BEHAVIOR and real-robot scores
Original title: Toward the Era of Agentic Embodied Intelligence
Authors report 0.6411 Q-score on 100 public tasks and 57.92% success on 12 real-robot tasks versus 26.67% for pi0.5.
The paper presents A2-MEAS, an embodied multi-agent framework that uses learned correction codes with MetaBrain supervision. Its action model, A2-WLA, adds motion-aware history and a latent world model to a vision-language-action policy. On the 50-task BEHAVIOR 2026 public subset it reports 0.5755 Q-score before online correction. The authors say they will release the system, models, training scripts, and infrastructure.
Unconfirmed: Scores are author-reported; the 2.21x figure spans different benchmark editions and training setups.
- On all 100 BEHAVIOR 2026 public tasks, A2-MEAS reports 0.6411 Q-score and 45.40% success.
- On RoboTwin 2.0, A2-WLA reaches 93.98% average success, 95.05% with the full system.
- Training infrastructure delivers 427.1 samples per second, 4.03x a pi0.5 reference.
- A history adapter lifts two pour-over recovery rates from 14% to 94% and 8% to 82%.
From the original article
1 min read at huggingface.co
Foundation models increasingly guide robot actions, yet sustained autonomy requires efficient control and failure recovery. We introduce A2-MEAS, an embodied multi-agent framework built around code as a scaling interface. Learned correction codes combine corrective language, spatial evidence, execution history, and action proposals.
By Chao Liang, Ke Chen, Fengxiaoxiao Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.