Stanford RL method fixes VLA latency, lifting robot success from 42% to 97% with 10 minutes of data
burny_tech · x · 2026-09-21
A Stanford team (Chelsea Finn, Dorsa Sadigh et al.) published "Reinforcement Learning for Real-Time Vision-Language-Action Policies," tackling the core bottleneck of using large VLA models for reactive robot control: high inference latency means observations are stale by execution time, causing distribution shift and degraded reliability.
Approach
- Decouple slow, expressive action generation from fast, reactive editing: the large pretrained VLA proposes action chunks in the background, while a lightweight RL edit policy reacts to state changes using the latest observation.
- Built on the EXPO-FT RL fine-tuning framework, instantiated as Real-Time EXPO-FT — the first to bring RL fine-tuning up to real-time control requirements for dynamic real-world manipulation.
Results
- On the Kinetix benchmark, the delayed policy beat all non-delayed methods in 10/10 environments.
- On four dynamic real-world tasks (object passing, ball balancing, table soccer kicking, dynamic picking), with online robot data capped at 10 minutes, average success improved from 42% to 97%.
More from Embodied
- UBTECH sells 400 U1 Pro humanoids to Yongda Auto in ~$10M deal — CyberRobooo · 2026-11-12
- Baidu World 2026 to showcase Apollo Go robotaxi rides and Apollo World physical AI model — Baidu_Inc · 2026-11-12
- NVIDIA's open-source Alpamayo 2 Edge VLA driving model lands in Autoware — drmapavone · 2026-11-12
- A closer look at Astra's robot gripper design (and a joke at its expense) — _Stocko_ · 2026-11-12
- Robot startup passes first presale milestone with 10+ units claimed — R4nsack · 2026-11-12
- Runway launches Praxis-1, a 27B MoE world action model, with a 73ms Fast variant — ChongZzZhang · 2026-11-12