Toward the Era of Agentic Embodied Intelligence

Toward the Era of Agentic Embodied Intelligence. By Chao Liang and nine others, from Shanghai Jiao Tong University and Anyverse Dynamics. Published on the sixth of November.

Foundation models increasingly decide what robots do. But keeping a robot autonomous over a sustained stretch asks for two more things: control that is efficient, and the ability to recover when something fails.

A2-MEAS is an embodied multi-agent framework built around one organising choice: code as the scaling interface. Its central object is a learned correction code, which bundles together corrective language, spatial evidence, the history of execution, and proposed actions. A component called MetaBrain supplies those codes in two roles. During training it provides supervision, and at runtime it provides guidance.

The knowledge moves in two ways. In training, a correction encoder and an adapter absorb it into the policy's weights. At deployment, the weights stay frozen and fresh codes adapt the behavior instead. The action model itself, A2-WLA, takes a vision-language-action policy and adds two things: a motion-aware history and a latent world model.

The authors call this design simple and say its strength was a surprise. With action pretraining and task specialization, and before any online correction, it reaches a Q-score of zero point five seven five five on the fifty-task public subset of BEHAVIOR two thousand twenty-six. That is two point two one times the zero point two six zero five reported by the two thousand twenty-five winner, though the abstract notes the comparison crosses benchmark editions and training configurations.

On all one hundred public tasks, the full A2-MEAS reaches a Q-score of zero point six four one one and a task success of forty-five point four percent. On RoboTwin two point zero, A2-WLA alone averages ninety-three point nine eight percent success, and the complete system lifts that to ninety-five point zero five percent.

On real robots, across twelve tasks in five household settings, A2-WLA averages fifty-seven point nine two percent success. The comparison, pi zero point five, averages twenty-six point six seven percent. In a long-horizon task, preparing pour-over coffee, a history-conditioned adapter changes what happens after a failure at two filter-handling steps. Compared with A2-WLA simply retrying, completion at the first step rises from fourteen percent to ninety-four. At the second it rises from eight percent to eighty-two.

The training infrastructure is measured too. It delivers four hundred and twenty-seven point one samples per second, which is four point zero three times the throughput of a pi zero point five reference implementation on matched hardware.

The thread through all of this is a single interface doing two jobs. The same correction codes that get folded into the weights while training are the ones that steer a frozen policy once it is deployed, so behavior can change in the field while the weights stay fixed. The authors say they will release the system, the models, the training scripts and the infrastructure.