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.
Figures & tables
Figure 1: The A 2 -MEAS framework. The MetaBrain coordinates task reasoning and recovery. Grounding agents provide specialized evidence. The correction interface combines recovery language, spatial evidence, executed history, and a fresh base-action proposal to modulate A 2 -WLA. These agents cooperate to control one robot.
Agents
Responsibility
Function
Task verification
Maintain the task goal and verify execution progress
Failure diagnosis
Review attempted actions and request missing evidence
MetaBrain
Recovery instruction
Describe how the next attempt should change
Perceptual grounding
Identify objects, regions and visible properties
Spatiotemporal grounding
Track identities, relative positions and changes over time
Diagnostic agents
Interaction and feasibility
Assess contact, support and movement constraints
Table 1: The three groups of agents in A 2 -MEAS. Rows within each group describe its responsibilities or internal components. Diagnostic agents can share tools across functions. A common execution layer dispatches controls, checks outcomes and records the history used by all three groups.
Figure 2: State progression during the closed-loop recovery of a misaligned filter placement. Following a loss of contact support along one edge of the support ring, the execution layer pauses to verify spatial relationships and synthesizes a sequence of lifting, realigning, and reseating actions conditioned on grounded high-level guidance. Faded outlines trace the initial misaligned pose, orange arrows indicate the directional motion flow, vertical dashed lines mark the central axis of the cup opening, and green indicators highlight the restored planar support.
Table 2: Structure of the correction context provided to the model to guide recovery. It combines the task objective, failure evidence, executed action history, and corrective guidance. Spatial evidence includes camera, frame, and coordinate references. A^tbase denotes a fresh, uncorrected action proposal generated from the current recovery state.
Figure 3: The correction interface within action generation. Language, spatial evidence, executed history and a fresh base proposal are encoded once per candidate. At every flow step, the adapter uses the current action hidden states to produce position-dependent residuals. Injection after the ninth of 18 action layers allows subsequent layers to convert the modified state into a corrected flow velocity. Orange modules are trained during correction learning while the base model is frozen. All parameters remain fixed during deployment.
Figure 4: Radio activation (Task 000). (a) Task context. (b–c) Adapter-guided motion (orange), base-policy continuation (teal), and success. (d–e) A separate A 2 -WLA-only rollout misses the circled button. Trajectories project left-gripper motion. The branches show different rollout instances, with panel (a) providing context rather than a shared initial state.
Figure 5: Offline construction of correction supervision. A frozen policy predicts a candidate at a demonstration state. Candidate and expert actions are aligned in time, units and coordinates, and windows with a large deviation at critical manipulation steps are retained. Visual evidence and the candidate-to-expert direction support context construction. The expert continuation is a training target. The projected positions in the example are an offline quality check.
Figure 6: Execution and recovery as one feedback loop. Ordinary action generation records the executed prefix and its observed outcome. An abnormal event invokes the MetaBrain and the required grounding agents. A fresh recovery state conditions both the base proposal and corrected generation. Failed attempts update the evidence and consume the remaining event budget. Verified recovery resumes execution toward the unfinished task goal. A timeout, exhausted budget or failed feasibility check exits recovery.
Figure 7: Architecture of A 2 -WLA. A motion-aware history encoder converts six-frame multiview histories into visual tokens. A 2B Qwen3-VL backbone combines these tokens with the task instruction to produce current context and a latent action. The world model predicts future visual features, which a spatial compressor aggregates into future context. An 18-layer action expert uses both contexts to generate a 50-step action chunk through ten-step flow integration, while a separate head estimates per-step progress. During recovery, a correction adapter injects a code-conditioned residual after the ninth layer. The lower panel shows how adjacent differences between paired-frame representations provide motion features for gated cross-attention and residual fusion before visual token merging.
Setting
World model
VLM
Continuous action
Global batch
4,608
2,048
2,048
Per-GPU batch
72
32
32
Training plan
Up to 10 epochs
150,000 steps
500,000 steps
Save interval
5,000 steps
20,000 steps
20,000 steps
Initial learning rate
3×10−4
7.5×10−5
7.5×10−5
Warm-up
10,000 steps
2,000 steps
2,000 steps
Table 3: Recorded schedules for the three base-model stages. The correction schedule is given in Section 2.5 . Training configuration and systems throughput measurements have different hardware scales and are reported separately.
Figure 8: Data processing and distributed execution performance before and after optimization. (a) single-GPU data processing time, (b) single-GPU step time, (c) eight-GPU step time, (d) eight-GPU throughput, and (e) eight-GPU time to first step. Approximate values are marked with ∼ . Percentages below the panels indicate relative decreases ( ↓ ) or increases ( ↑ ), calculated from the reported values. Single-GPU and eight-GPU comparisons use their respective fixed configurations and do not form one continuous ablation sequence.
Figure 9: Throughput gains from (a) dedicated computation paths and scheduling, (b) static token, rotary-index and metadata caches, and (c) Sharded Local Muon. Each panel compares its own fixed configuration. Gains from different panels cannot be added.
Setting
Naive eight-GPU baseline
Optimized evaluation
Worker-pool GPUs
8 (projected)
8 Pro6000 GPUs
Workers per GPU
1
3
Concurrent workers
8
24
Evaluation episodes
1,000
1,000
Wall-clock duration
≈159.4 hours
18.1 hours
Episode throughput
≈6.27 per hour
≈55.2 per hour
Table 4: Evaluation runtime for 100 BEHAVIOR 2026 tasks and 1,000 episodes. The naive eight-GPU baseline is projected as 1,275/8≈159.4 hours under ideal linear scaling of the estimated serial-worker runtime. The optimized result is measured. GPU counts cover simulation and policy workers, with the reasoning service hosted separately.
Figure 10: Distribution of demonstration durations in the BEHAVIOR 2026 dataset. The dashed lines indicate the mean (5.86 minutes) and median (5.58 minutes). Reproduced from the official dataset statistics [ 26 ] .
Configuration
Q-score
Success (%)
Successes
Coverage
A 2 -WLA single model
0.4597
27.20
272/1,000
70/100
A 2 -WLA multiple models
0.6040
40.60
406/1,000
80/100
A 2 -MEAS complete system
0.6411
45.40
454/1,000
88/100
Table 5: Public-set evaluation results for BEHAVIOR 2026, with 100 tasks and 1,000 recorded episodes per configuration. Coverage counts tasks with at least one complete success.
Figure 11: Task-level performance in the public-set evaluation records, ordered by decreasing mean Q-score, with ties resolved by task ID. Blue dots show Q-score, with labels rounded to three decimals for display. The final column reports complete successes out of ten recorded instances. The two panels show ranks 1–50 and 51–100. Task names are shortened for display. IDs identify the benchmark tasks.
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
Table 6: Public-set results for BEHAVIOR 2026. We include our self-reported result and the 15 highest-scoring entries in the published self-report list, ordered by increasing Q-score. These records follow the participating teams’ reported protocols. Institution names follow leaderboard and archived participant records. Scores were accessed 31 October 2026 [ 29 , 30 ] .
Method
Q-score
SR (%)
Evaluation context
Competition public-dataset evaluation [ 17 ]
RLC [ 18 ]
0.2605
11.20
4 checkpoints
Comet [ 17 ]
0.1830
14.40
Competition submission
SimpleAI Robot [ 17 ]
0.1943
14.00
Competition submission
The North Star [ 17 ]
0.1702
12.80
Competition submission
Post-competition and subsequent reports
Table 7: Public-set results on task IDs 0–49 of BEHAVIOR 2026, with 500 recorded episodes per configuration, and BEHAVIOR 2025 public-dataset results grouped by evaluation setting. A dash denotes an unreported success rate.
Configuration
History frames
Latent world model
FAST init.
Macro SR (%)
π0.5 (baseline)
—
—
—
26.67
A 2 -WLA configurations
Config. 1 (base fine-tuning)
2
×
×
39.58
Config. 2 (+ latent world model)
2
✓
×
47.08
Config. 3 (+ six-frame history)
6
✓
×
54.17
A 2 -WLA (full configuration)
6
✓
✓
57.92
Table 9: Base-policy configurations and results. “+” highlights an added component. ✓ and × denote enabled and disabled options. “—” in the settings columns denotes an unrecorded setting. FAST initialization applies to the vision-language backbone and history encoder. Evaluation conditions and qualifications are given in Table 10 .
Figure 12: Execution examples and results across 12 real-world tasks. Panels follow the task order in Table 10 and show keyframes from A 2 -WLA demonstration videos, labeled with the visible actions. The bars below each frame compare successes / attempts for π0.5 and the full A 2 -WLA configuration, with 20 trials per task and a common success-rate scale of 0–100%.
Figure 13: Grasp correction and step completion during pour-over coffee preparation. (a) Rack pickup in Step 2. (b) cup-rim pickup in Step 7. Each row shows a missed grasp, gripper opening and adjustment, regrasping, and lifting. Timestamps are playback positions in the respective clips. Stages are labeled from visible actions. (c) Base-policy execution with and without the Adapter from the original task start. (d) base-policy retry and the failure-history-conditioned Adapter after rollback. Each measured condition has 100 trials per step. Measured bars show completions / attempts.
Setting
Model components
Training data
Performance
World model
Motion encoding
Skill sampling
Successful rollouts
Q-score
SR (%)
Observation and initialization baselines
Current frame only
–
–
–
–
0.1582
8.33
Six-frame history
–
–
–
–
0.2167
13.33
FAST pretraining
–
–
–
–
0.3453
20.83
Action-policy initialization
–
–
–
–
0.4846
39.72
Table 11: Base-model ablations on Selected12, averaged over three evaluation rounds of 120 episodes each. The upper block compares observation inputs and initialization. The lower block adds model and data components cumulatively to the action-pretrained baseline. Check marks denote enabled components and dashes denote their absence. History-input configurations retain conventional history encoding when motion-aware encoding is disabled.
Model
Condition change
Mean Q
Success (%)
A 2 -WLA (w/o Adapter)
No Adapter
0.5310
41.67
Adapter w/o Language & Grounding
Remove correction language, task text and structured grounding
0.5344
43.33
Adapter w/o Long History
Retain only the latest state–action record
0.5356
44.17
Adapter w/o Candidate
Remove base-action candidate encoding
0.5542
47.50
Adapter w/o Grounding
Remove structured grounding input
0.5570
49.17
Full Adapter
All conditions retained
0.6108
52.50
Table 12: Adapter condition ablation on the 12-task correction subset, with 120 episodes per configuration. All variants share the A 2 -WLA weights.
Interface
Q-score
Physical recovery rate
Effective continuation rate
Final task completion rate
Instruction restatement (frozen A 2 -WLA)
0.5668
46.27%
25.37%
38.81%
End-effector geometric correction (VLM + IK)
0.6083
52.05%
28.77%
45.21%
Latent-space residual adapter
0.6108
54.93%
33.80%
43.66%
Table 13: Recovery and continuation with three correction interfaces on the 12-task Adapter evaluation subset.
Figure 14: Adapter generalization on the last ten instances of each of the 12 correction tasks, with 120 episodes per configuration. Both axes start at zero.
Figure 15: Local correction and complete-episode outcomes for hanging pictures, bringing water, and washing a baseball cap. Blue trajectories show A 2 -WLA-only execution. Orange trajectories show Adapter-controlled intervals. Small photographic insets show subsequent observations from failed A 2 -WLA-only rollouts in the upper row and terminal observations from successful Adapter-enabled rollouts in the lower row. Colored boxes magnify the local trajectories. Each pair comprises independently executed rollouts on the same task instance.
Task
Task name
B: Q
F: Q
B: success
F: success
000
Turn on radio
0.7000
0.7000
7/10
7/10
001
Pick up trash
0.6000
0.4667
5/10
3/10
006
Hide Easter eggs
0.3667
0.3778
3/10
3/10
017
Bring water
0.8333
0.3000
8/10
3/10
018
Tidy bedroom
0.7000
0.7333
5/10
3/10
022
Put shoes on rack
0.8500
0.7900
7/10
5/10
Table 14: Complete-task conditioning (B) and VLM-selected Focus (F) on Selected12, both trained for 180k steps. Each configuration reports one round of ten instances per task. The baseline results are from its first evaluation round.
Figure 16: Task-level changes from complete-task conditioning to Focus on Selected12. The left panel shows Q-score differences and the right panel shows success-rate differences in percentage points.
Mode
Conditioning and scheduling
Q-score
SR (%)
F-VLM
Full task and online VLM-selected Focus
0.4411
30.83
F-Global
Full task only
0.4201
30.00
F-Linear
Simulator-assisted assessment, fixed order
0.4115
25.00
F-Adaptive
Same assessment, reordering and revisitation
0.3785
27.50
Table 15: Scheduling ablation with one frozen Focus action model and 120 scheduled episodes per mode. Q-score is reported on a 0–1 scale. Missing or failed evaluation runs are assigned zero Q-score and counted as unsuccessful. Valid incomplete trajectories retain their measured partial Q-score.
Figure 17: Workflow variation and sparse transition histories. Common variation covers 75 of 100 tasks. The 30 drill-placement demonstrations cover 16 preceding-history combinations, with 1–5 examples per combination.
Figure 18: Base-policy recovery without the Adapter, beyond the corresponding task demonstrations. Bottle reorientation and corn regrasping occur in episodes that ultimately time out after prolonged retries. Glove retrieval is followed by continued washing and task completion. Each row follows one episode chronologically.
Figure 19: Physical state changes during failed task execution. (a) During garden furniture transport, a cart tips over and subsequent manipulation fails to right it. (b) During room rearrangement, a bedside cabinet tips over and the lamp falls into a gap between furniture. (c) During food canning, cut food pieces scatter into the sink and remain scattered until the episode ends. Columns show the initial conditions, state changes, and subsequent observations within the same episode. Each row represents an independent failed episode. Timestamps indicate playback positions in the source videos; images show cropped head-camera views.