SYSTEM: Distilling Foundation Model Capabilities into Action Policies
MetaBrain maintains the goal, diagnostic agents resolve physical questions, and a grounded correction interface guides A2-WLA. Nominal control repeats action and observation; diagnosis and correction are invoked when recovery is needed.
MetaBrain’s recovery instruction, verified spatial evidence, executed history and a fresh base-policy proposal are encoded into a correction code. A residual adapter uses this code to adjust action generation, with all model parameters fixed during deployment.
Correction examples are mined offline by comparing a frozen policy’s unexecuted proposals with expert demonstrations. Grounded visual evidence and corrective instructions form the training context; the expert continuation supervises the correction branch while the base policy stays frozen.
When execution becomes abnormal, the system pauses pending controls and gathers fresh evidence for a new recovery attempt. It checks the corrected candidate before execution, verifies the outcome and either resumes the unfinished task or updates the recovery query within the remaining budget.
MODEL: Simple but Surprisingly Strong World-Language-Action Model
A2-WLA combines recent visual history with predicted scene features to generate continuous action chunks. Its history encoder captures observed change, while the latent world model supplies future context alongside the robot state.
INFRASTRUCTURE: 4× Training Throughput of π0.5 at a Similar Model Size
The training pipeline reduces repeated data preparation, reuses shared model computation, and distributes optimizer work across GPUs.
Data processing and distributed execution
Batch JPEG decoding, Apache Arrow caches and asynchronous prefetching reduce repeated input work, cutting data preparation from about 1,300 ms to 28 ms and single-GPU step time from 3.03 s to 1.13 s. Parallel dataset preparation and streamlined synchronization improve multi-GPU startup and execution.
Training-throughput optimizations
Dedicated computation paths and scheduling, static caches and Sharded Local Muon yield independent throughput gains of 36.97%, 5.70% and 1.7%, respectively. With shared-prefix execution and the other pipeline optimizations, the final eight-GPU implementation reaches 427.1 demonstration samples/s, versus 106 for the π0.5 reference—a 4.03× gain.
SIMULATION
Swipe to see all columns
| Configuration | Q-score | Success (%) | Successes | Coverage |
|---|---|---|---|---|
| A2-WLA single model | 0.4597 | 27.20 | 272/1,000 | 70/100 |
| A2-WLA multiple models | 0.6040 | 40.60 | 406/1,000 | 80/100 |
| A2-MEAS complete system | 0.6411 | 45.40 | 454/1,000 | 88/100 |
Swipe to see all columns
| Team or submission | Institution | Q-score | SR (%) |
|---|---|---|---|
| George Albrecht | CAU / GMU / EWHA | 0.3862 | 20.00 |
| Mirua | Xiaomi Robotics | 0.3971 | 19.80 |
| VERA | AXXX | 0.4000 | 25.00 |
| Newbie | morphic & 4dvlab | 0.4040 | 21.70 |
| Huawei AI Algorithm Dept | Huawei Technologies | 0.4076 | 25.20 |
| Lion | Lionrock AI Lab | 0.4082 | 21.80 |
| The Southern Cross | Independent | 0.4099 | 23.00 |
| MXI | Beijing 1011 Technology | 0.4187 | 22.70 |
| Sora | Independent | 0.4195 | 22.50 |
| Multi | mor | 0.4302 | 19.80 |
| Anyverse A2R | Anyverse-A2R | 0.4321 | 24.70 |
| VGI-SKT | SNU / SKT | 0.4449 | 22.30 |
| Robot Learning Collective | Independent | 0.4458 | 27.30 |
| tasman3001 | Ambarella, Inc. | 0.4808 | 30.40 |
| asteroid | CLab | 0.5143 | 30.80 |
| A2-MEAS (ours) | Anyverse–AutoLab | 0.6411 | 45.40 |
Swipe to see all columns
| Method | Q-score | SR (%) | Evaluation context |
|---|---|---|---|
| Competition public-dataset evaluation [11] | |||
| RLC | 0.2605 | 11.20 | 4 checkpoints |
| Comet | 0.1830 | 14.40 | Competition submission |
| SimpleAI Robot | 0.1943 | 14.00 | Competition submission |
| The North Star | 0.1702 | 12.80 | Competition submission |
| Post-competition and subsequent reports | |||
| Openpi Comet [23] | 0.3450 | — | 2 checkpoints, reinforcement fine-tuning |
| π0.5 [7] | 0.2626 | — | 4 epochs, single checkpoint |
| G0.5 [7] | 0.2904 | — | 1 epoch, single checkpoint |
| G0.5 [7] | 0.3136 | — | 4 epochs, single checkpoint |
| Our 2026 public-set subset, 50 tasks and 500 episodes | |||
| A2-WLA single model | 0.4522 | 25.60 | 4 epochs, single checkpoint |
| A2-WLA multiple models | 0.5755 | 37.40 | 25 checkpoints |
| A2-MEAS complete system | 0.6191 | 41.40 | Including the full system |
Swipe to see all columns
| Method or configuration | Clean (%) | Randomized (%) | Average (%) |
|---|---|---|---|
| π0.5 [9, 34] | 82.74 | 76.76 | 79.75 |
| LingBot-VLA (without depth) [34] | 86.50 | 85.34 | 85.92 |
| LingBot-VLA (with depth) [34] | 88.56 | 86.68 | 87.62 |
| Motus [35] | 88.70 | 87.00 | 87.85 |
| StarVLA [36] | 88.70 | 87.80 | 88.25 |
| LingBot-VA [37] | 92.90 | 91.50 | 92.20 |
| Fast-WAM [38] | 91.90 | 91.80 | 91.85 |
| G0.5 [15] | 93.70 | 92.80 | 93.25 |
| LingBot-VLA 2.0 [39] | 93.52 | 92.80 | 93.16 |
| MotuBrain [40] | 95.80 | 96.10 | 95.95 |
| Our configurations | |||
| A2-WLA | 94.26 | 93.70 | 93.98 |
| A2-MEAS | 95.08 | 95.02 | 95.05 |
REAL WORLD
Following a missed filter grasp, the robot opens and adjusts the gripper, regrasps, and lifts the filter clear to continue the task. From the original start, the Adapter raises step completion from 71% to 98% in Step 2 and 76% to 95% in Step 7.
After failure and rollback, the failure-history-conditioned Adapter achieves 94% and 82% step completion, compared with 14% and 8% for base-policy retry, respectively. Each condition uses 100 trials per step.
Click a video to enlarge.
Loading task inventory…
Across 12 tabletop tasks in five household settings—study, bedroom, living room, kitchen, and entryway—A2-WLA is evaluated on grasping, placement, action sequencing, and contact control.
With high-level feedback and test-time correction disabled, the full A2-WLA configuration achieves 57.92% macro-averaged whole-task success over 20 trials per task, compared with 26.67% for π0.5.
THE PEOPLE BEHIND THE WORK
AutoLab, School of Artificial Intelligence, Shanghai Jiao Tong University
Research Lab, Anyverse Dynamics
@misc{liang2026eraagenticembodiedintelligence,
title = {Toward the Era of Agentic Embodied Intelligence},
author = {Chao Liang and Ke Chen and Fengxiaoxiao Li and
Shibo Huang and Yufeng Sun and Zhaokai Yin and
Jianhuan Lai and Haozhe Feng and Chenghuan Jia and
Zhipeng Zhang},
year = {2026},
eprint = {2611.09124},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2611.09124}
}