Real-Time EXPO-FT: decoupling slow VLA brains from fast reactive edits

Stanford researchers pair slow pretrained VLA proposals with high-frequency RL action edits to solve the latency lag on real robots, jumping from 42% to 97% success in 10 minutes.

Primary source: arXiv:2609.18207: https://arxiv.org/abs/2609.18207

What’s new: In Real-Time EXPO-FT (arXiv:2609.18207), Stanford researchers Perry Dong, Kuo-Han Hung, Dorsa Sadigh, and Chelsea Finn target the latency bottleneck in robot learning. Large Vision-Language-Action (VLA) models offer broad semantic priors, but their forward passes take hundreds of milliseconds. By the time a multi-step action chunk reaches physical actuators, the camera frame that generated it is already stale.

Instead of running slow end-to-end inference or relying on passive imitation learning, Real-Time EXPO-FT splits control into two frequencies:

1. The slow brain: A large pretrained VLA periodically proposes action trajectories using its broad behavioral and semantic priors.
2. The fast brain: A lightweight edit policy runs at high frequency, ingesting the newest camera frame and editing the pending action chunk in real time.

The team formulates reinforcement learning fine-tuning directly over this delayed Markov Decision Process. On the Kinetix physics benchmark, Real-Time EXPO-FT outperformed all delayed and non-delayed baselines across 10 out of 10 environments. In physical experiments spanning dynamic object passing, ball balancing, table soccer kicking, and dynamic picking, autonomous online RL fine-tuning with only 10 minutes of robot interaction improved policy success from 42% to 97% without human intervention.

Why it matters: Foundation models for robotics are useless if inference latency makes closed-loop control lag behind reality. Decoupling semantic trajectory proposals from high-frequency reactive edits lets robots keep the reasoning power of large models while reacting to physical perturbations (such as human nudges or rolling balls) at hardware speed. Demonstrating 97% success after 10 minutes of autonomous fine-tuning suggests sample-efficient RL can bridge the gap between static pretraining and live physical contact.

Caveats: Preprint. The fast policy operates as a residual editor; if the underlying VLA produces a completely misguided semantic plan, local action corrections cannot rescue the task. Physical trials focused on short-horizon dynamic motor skills (balancing, kicking, picking) rather than long multi-room mobile manipulation. The system requires an accurate streaming pipeline and synchronized actuation buffers to compute delays reliably.