Driving · 2026-11-09

Toward the Era of Agentic Embodied Intelligence

Chao Liang, Ke Chen, Fengxiaoxiao Li, Shibo Huang, Yufeng Sun, Zhaokai Yin, Jianhuan Lai, Haozhe Feng, Chenghuan Jia, Zhipeng Zhang

arXiv:2611.09124PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors introduce A2-MEAS, an embodied multi-agent framework using “code as a scaling interface” for corrective behavior. MetaBrain generates recovery supervision and guidance; diagnostic agents provide grounded evidence; a learned correction encoder and adapter inject correction codes into a vision-language-action policy (A2-WLA) during training, while deployment keeps A2-WLA backbone weights fixed. The system reports strong BEHAVIOR 2026 and real-robot recovery results and releases code and models.

Why it matters

The paper targets sustained embodied autonomy by combining learned control with evidence-based failure recovery. By encoding corrective instructions plus grounded spatial evidence and execution history into a “correction code” that modulates a continuous-action policy without updating its weights at deployment, the authors aim to make foundation-model reasoning usable for long-horizon control and online recovery.

Method

  • A2-MEAS uses code as a scaling interface: learned correction codes combine corrective language, grounded spatial evidence, executed history, and a fresh action proposal to modulate an action expert via a residual adapter.
  • MetaBrain provides corrective supervision during training and runtime guidance, while diagnostic agents supply perceptual, spatiotemporal, and interaction evidence using a shared execution record.
  • The action model A2-WLA augments a vision-language-action policy with motion-aware history and a latent world model predicting future visual features; deployed recovery uses fresh codes with all A2-WLA parameters fixed.

Limitation

The current correction process pauses execution for diagnosis, which limits its use for events that demand immediate reactions.

Abstract (from arXiv)

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. MetaBrain provides training supervision and runtime guidance. A correction encoder and adapter absorb this knowledge into policy weights during training, while fresh codes adapt behavior with frozen weights at deployment. The action model, A2-WLA, adds motion-aware history and a latent world model to a vision-language-action policy. Surprisingly, this simple design with action pretraining and task specialization reaches 0.5755 Q-score on the 50-task BEHAVIOR 2026 public subset before online correction, 2.21x the 2025 winner's reported 0.2605 across different benchmark editions and training configurations. On all 100 BEHAVIOR 2026 public tasks, A2-MEAS achieves 0.6411 Q-score and 45.40% task success. On RoboTwin 2.0, A2-WLA achieves 93.98% average success, rising to 95.05% with the complete system. Across 12 real-robot tasks in five household settings, A2-WLA achieves 57.92% average success versus 26.67% for pi0.5. In long-horizon pour-over coffee preparation, a history-conditioned adapter raises post-failure completion at two filter-handling steps from 14% to 94% and from 8% to 82%, compared with A2-WLA retries. Our training infrastructure delivers 427.1 samples per second, 4.03x the throughput of a pi0.5 reference implementation on matched hardware. We will release the system, models, training scripts, and infrastructure to support efficient embodied learning.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First announced by arXiv: 2026-11-09 · Indexed by Pub-AI: 2026-11-09 UTC

preprint

Toward the Era of Agentic Embodied Intelligence

Chao Liang, Ke Chen, Fengxiaoxiao Li, Shibo Huang, Yufeng Sun, Zhaokai Yin, Jianhuan Lai, Haozhe Feng, Chenghuan Jia, Zhipeng Zhang

First announced by arXiv: 2026-11-09 · Original record ↗

arxiv
2611.09124
Explicit source versions (1)
  • v1 · Date not supplied
  1. Preprint available

    Source date: 2026-11-09 · First detected: 2026-11-09 UTC · Historical baseline

    Source evidence (1)
    • Official evidence ↗ · RSS guid=oai:arXiv.org:2611.09124v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv

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. MetaBrain provides training supervision and runtime guidance. A correction encoder and adapter absorb this knowledge into policy weights during training, while fresh codes adapt behavior with frozen weights at deployment. The action model, A2-WLA, adds motion-aware history and a latent world model to a vision-language-action policy. Surprisingly, this simple design with action pretraining and task specialization reaches 0.5755 Q-score on the 50-task BEHAVIOR 2026 public subset before online correction, 2.21x the 2025 winner's reported 0.2605 across different benchmark editions and training configurations. On all 100 BEHAVIOR 2026 public tasks, A2-MEAS achieves 0.6411 Q-score and 45.40% task success. On RoboTwin 2.0, A2-WLA achieves 93.98% average success, rising to 95.05% with the complete system. Across 12 real-robot tasks in five household settings, A2-WLA achieves 57.92% average success versus 26.67% for pi0.5. In long-horizon pour-over coffee preparation, a history-conditioned adapter raises post-failure completion at two filter-handling steps from 14% to 94% and from 8% to 82%, compared with A2-WLA retries. Our training infrastructure delivers 427.1 samples per second, 4.03x the throughput of a pi0.5 reference implementation on matched hardware. We will release the system, models, training scripts, and infrastructure to support efficient embodied learning.