Skip to main content
Back to timeline
arXivSource publication:

Astra stops being called every action chunk: a three-tier system reaches 79.13% success on RoboMME with 3.63 calls per episode

Synopsis

Robotic manipulation often requires remembering information that is no longer visible, and calling a large vision-language planner before every action chunk is slow and costly. A three-tier system separates the GPT-6 Astra planner, a vision-language-action policy, and a fine-tuned Qwen3-VL-4B monitor: Astra is called at episode start and again only when the monitor reports subtask completion or a task-specific event occurs. On the official RoboMME test set, the system reaches 79.13% overall success with an average of 3.63 Astra calls and 80.84 seconds per episode, 34.62 points above the best published non-oracle baseline and 4.95 points below GroundSG+Oracle, which supplies ground-truth subgoals.

Source-provided article image: Can Astra Solve RoboMME without Breaking the Bank?
arXiv

Interpretation

The system replaces per-action-chunk planner calls with event-triggered calls and reaches 79.13% overall success on RoboMME. Previously a large vision-language planner was called before every action chunk, which is slow and mostly redundant; the best published non-oracle baseline, FrameSamp+Modul, reaches 44.51% overall success. On the official RoboMME test set with 50 episodes per task and a 1,300-step budget per episode, the system reaches 79.13% overall success, 34.62 points above FrameSamp+Modul and 4.95 points below GroundSG+Oracle.

Astra selects one executable grounded subtask at episode start, the vision-language-action policy keeps executing the same instruction from new observations, and the fine-tuned Qwen3-VL-4B monitor decides after each action chunk whether the subtask is finished. The monitor makes a binary decision from eight recent front-camera frames, the front image from when the current instruction was issued, and the current wrist image; a negative decision keeps the current instruction for another chunk. The monitor is fine-tuned with LoRA on 71,835 examples built from RoboMME demonstrations with the vision encoder frozen; each action chunk runs 16 environment steps, and recorded trajectories contain 24.62 action chunks on average.

Perspective

Evaluation is limited to the official RoboMME test set with 50 episodes per task and a 1,300-step budget per episode, and only the full three-tier system is evaluated; the authors state that ablations are needed to separate the contributions of memory, planner capability and monitoring.

The authors have not yet attributed failures to planning, monitoring or execution errors, and they did not measure the monitor's closed-loop error rate; the simulator pauses while Astra reasons, so scenes that change during reasoning have not been tested.