EA-VLA: Execution-Aware VLAs for Robust Manipulation

The execution gap. A naïve VLA assumes its commanded action âẓ = âẽ is tracked exactly. In practice, tracking error, latency and actuator effects make executed motion diverge. EA-VLA conditions on execution feedback ê = âẓ − âẽ and produces compensated actions.

TL;DR

Actions predicted ≠ actions executed

Most VLAs treat the low-level execution interface as fixed and reliable. Real robots don't: tracking error, latency, actuator saturation and calibration drift open an execution gap that quietly breaks contact-rich tasks. EA-VLA augments policy inference with execution-state feedback and adapts pretrained VLAs through RL over randomized controller dynamics in simulation — learning to compensate for that gap.

Method

Execution-aware RL fine-tuning

A pretrained VLM + action expert is conditioned on robot state and execution feedback ê = âẓ − âẽ, then fine-tuned with RL in ManiSkill under domain-randomized controller dynamics — exposing the policy to a broad range of execution gaps without real-robot data.

Results

Robust across the execution gap

Success rate vs. execution-gap magnitude (randomization ratio σ). EA-VLA (green) degrades gracefully where the baseline collapses — in both simulation (dotted) and on the real robot (solid).

ManiSkill simulation — success rate across randomization ratio σ

Model	σ=0.0	σ=0.001	σ=0.005	σ=0.01	σ=0.02	σ=0.05
Supervised baseline	48.7%	49.7%	14.2%	4.3%	2.9%	2.5%
RL, no execution gap	83.4%	51.1%	11.2%	8.1%	4.4%	2.2%
EA-VLA (Ours)	77.5%	80.1%	60.4%	28.6%	6.1%	2.6%

Baselines overfit to nominal dynamics and collapse under minor drift; EA-VLA stays robust across two orders of magnitude more perturbation.

Zero-shot sim-to-real (WidowX)

Policy	Low gap	High gap
Baseline	0.63	0.47
EA-VLA	0.66	0.58

Under true proprioceptive feedback (high gap), the baseline degrades sharply while EA-VLA holds.

Ablation — state-history length H

H	σ=0.0	σ=0.001	σ=0.005	σ=0.01
0 (none)	83.4%	51.1%	11.2%	8.1%
1	70.3%	70.5%	51.7%	22.2%
2 (default)	77.5%	80.1%	60.4%	28.6%
4	76.3%	76.3%	60.3%	26.4%

Execution awareness drives the gains; H=2 balances robustness and cost.