ARLI: Latency-Aware RL Finetuning Fixes VLA Robot Policies Where Standard RL Fails

oier_mees · x · 2026-09-22

Researchers from Siemens, UC Berkeley, Microsoft, and ETH Zurich published Learning to Act While Waiting, tackling RL finetuning of generalist robot policies (VLAs) under inference latency.

Problem: Severe VLA inference latency means actions are computed from stale observations, altering effective environment dynamics and breaking the Markov assumption RL relies on — standard RL algorithms fail completely.
Method: The latency-aware ARLI framework (Asynchronous RL with Intermediate Information) builds on asynchronous inference that interleaves action generation with execution, restoring near-Markovian structure via state augmentations that incorporate committed actions and a mid-inference observation.
Results: Across simulated and real-world manipulation tasks, ARLI enables effective finetuning under delay where standard RL fails entirely, matching or exceeding no-latency standard RL performance.

Authors include Sergey Levine, Oier Mees, and Andrew Wagenmaker.

Original post →

ARLI: Latency-Aware RL Finetuning Fixes VLA Robot Policies Where Standard RL Fails

More from Embodied