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
cs.RO
2026-11-06
A2-MEAS steers a frozen policy with a correction code. On 100 BEHAVIOR 2026 tasks: 0.6411 Q-score and 45.40% success. Real-robot A2-WLA scores 57.92% vs 26.67% for π0.5.
In a long household task, the next motion depends on the state the previous one left behind. A recovered grasp can leave an object in an orientation that blocks the later place. A language planner can decide to try again. It does not choose the approach angle, the alignment, or the gripper timing.
A VLA maps images and instructions to continuous controls, so recurring choices can live in the weights. Unfamiliar failures still need a larger model. At deployment those weights stay fixed, while a diagnosis still has to change the next stretch of motion. A2-MEAS is the interface between the two.
Three roles share one execution record. A2-WLA emits continuous actions. On the public set, MetaBrain is Qwen3.8-27B and writes the recovery instruction. The action backbone is a separate 2B Qwen3-VL. Diagnostics run on demand: DINO features, SAM masks, tracking, and contact checks. If the evidence is thin, the answer stays unknown.
The correction code is a conditioning vector, not a program. It packs the recovery instruction, verified spatial evidence, executed history, and the base policy's fresh proposal, which has not been executed. The vector is fixed for a whole action chunk. A residual adapter sits after layer 9 of the 18-layer action expert. The later layers turn the edited hidden state into a flow velocity. At every flow step the adapter rereads that position's hidden state, so one chunk can lift first and realign later. Bottleneck width is 256. The correction branch has fewer than 100M trainable parameters. The output projection starts at 0 and the channel scale α starts at 0.1, so the residual starts at zero and the base policy is unchanged.
Correction training freezes the base and updates only the adapter. On demonstration states the frozen policy proposes a chunk. Windows are kept when that proposal misses the expert at skill changes, pre-contact approach, or gripper events. The cap is 2% of the 1,950-hour corpus, 39 hours. Those windows are proposal errors. After deployment every weight stays fixed, and a new code changes the motion.
A2-WLA reads six frames and uses adjacent-frame differences as motion cues. Frame spacing is not fixed, so the differences record visible change rather than calibrated speed. A shared latent predicts future features across cameras. Future frames are absent online, and the backbone estimates the latent from the current observation. Control is 10-step flow matching: predict 50 steps, execute at most 32, across 23 whole-body dimensions. The base policy trains in three stages. Correction is a fourth stage. One anomaly allows at most three physical retries and 60 seconds. A coordinator call times out at 10 seconds, with at most two extra views.
| Comparison | Result |
| 50-task specialized policies, before correction, vs RLC 2025 public score | Q-score 0.5755 vs 0.2605 |
| 100-task full system vs multi-model A2-WLA | 0.6411, 45.40% vs 0.6040, 40.60% |
| RoboTwin 2.0, full system vs action model | 95.05% vs 93.98% |
| 12 real tasks, A2-WLA vs π0.5 | 57.92% vs 26.67% |
| Two filter steps, adapter vs retries of the same policy | 94/100 and 82/100 vs 14/100 and 8/100 |
On the 50-task subset, one shared model scores 0.4522 before any online correction. 0.5755 is about 2.21 times 0.2605, and the paper states that the comparison crosses benchmark editions and training setups. On all 100 tasks the multi-model action baseline is 0.6040 Q-score and 40.60% complete-task success. MetaBrain, diagnostics, and learned correction move that to 0.6411 and 45.40%. At least one successful episode appears on 88 tasks. Each task has 10 evaluation instances, so that bar is low.
The RoboTwin rates, 93.98% and 95.05%, average the Clean and Randomized conditions. The full system is 1.07 percentage points above the action model. The 12 real tasks cover five household settings, 20 trials per task and configuration. A2-WLA at 57.92% is about 2.17 times π0.5 at 26.67%, with no high-level correction. Pour-over coffee takes about four minutes. The filter comparison is the same frozen A2-WLA with the adapter switched off, and it supplements the correction tests in simulation.
Training throughput is 427.1 demonstration samples per second on eight GPUs at global batch 256. A π0.5 reference on the same hardware, image resolution, and action horizon reaches 106 samples per second, about 4.03 times lower throughput. Each implementation picks its own batch size to fill the machine. The 1,000-episode suite finishes in 18.1 hours on eight Pro6000 GPUs with 24 concurrent workers. An estimated 1,275-hour serial runtime, spread across eight GPUs with ideal linear scaling, is 159.4 hours, about 8.8 times slower. The reasoning service sits on separate machines.
Most of the public score sits in the action model. Before online correction, task-specialized policies already reach 0.5755 on the 50-task subset. On the 100-task average, correction adds about 0.0371 Q-score and 4.8 percentage points.
A missed filter seat is a different story. Retries from the same frozen policy finish 14/100 and 8/100 of those two steps. A history-conditioned adapter finishes 94/100 and 82/100. Recurring fixes go into a small adapter, the large model writes an instruction only when something looks wrong, and the base weights stay put. A full simulation suite in 18.1 hours makes a daily long-horizon loop practical. The paper says the system, models, training scripts, and infrastructure will be released.
Correction examples are gaps between unexecuted proposals and expert continuations. Deployment sees physical states after a real failure. The paper calls that a distribution gap. The mined offset is defined on controls. The adapter edits hidden-state residuals. The method sections do not isolate which geometric change a residual implements.
The 2.21 times figure is not a rank on one leaderboard. End-effector position and rotation losses do not check collision, contact stability, or dynamic feasibility. A progress score near 1 does not mean the task is finished.
The main 12-task real-robot result has no high-level correction. The coffee run compares a history-conditioned adapter with retries, so real long-horizon gains for the full coordinator still lack a number on the same footing as simulation. The 8.8 times evaluation speedup is against a projected baseline. Complete-task success is 45.40%. Most episodes still fail.