Siemens and Berkeley introduce ARLI for RL fine-tuning of VLAs under inference latency

svlevine · x · 2026-09-22

ARLI (Asynchronous RL with Intermediate Information) from Siemens, UC Berkeley and collaborators tackles why async VLA inference breaks the Markov assumption needed for RL fine-tuning. A small, faster RL policy observes more recent images and steers a large robot foundation model, using state augmentations (committed actions + mid-inference observation) to restore near-Markovian structure. Experiments show ARLI enables effective finetuning under latency where standard RL fails entirely, even matching no-latency RL performance.

Original post →

More from Embodied

Embodied channel →