Robotics · 2026-10-16
Fast-VLA: You Only Think Once for Robot Manipulation
Kehong Gong, Guanli Hou, Xiaoxu Zheng, Qi Wang, Mingxi Xu, Anbang Zhai, Xiaoyu He, Mingyuan Zhang, Michael Bi Mi
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
Fast-VLA invokes a frozen 4B VLM once per episode to encode the instruction and initial views into fixed, multi-layer KV task memory. A lightweight online actor then uses learned time- and arm-indexed queries with current visual/proprioceptive feedback to query this fixed memory and predict action chunks without re-running the VLM. The authors report strong success on RoboTwin 2.0 and LIBERO/LIBERO-Plus and lower RoboTwin amortized inference time versus continually conditioned baselines.
Why it matters
The authors aim to decouple task understanding (slow-changing) from closed-loop control (fast-changing) by establishing episode-level task context once and keeping it fixed during execution. This reduces repeated foundation-model calls in the feedback loop while retaining responsiveness via querying fixed multi-layer task memory at each action-chunk boundary.
Method
- Frozen episode-level memory: a frozen 4B VLM encodes instruction + initial multi-camera views into fixed, layer-wise key–value task memory once per episode.
- Online query actor: learned time- and arm-indexed queries fuse with current visual and proprioceptive feedback, then use cross-memory attention to retrieve fixed task context for action prediction.
- Action chunk prediction without re-running the VLM: the action head predicts a 50-step bimanual action chunk in one forward pass at each chunk boundary.
Limitation
Handling changes in task goals or newly introduced instructions during an episode remains unexplored, since Fast-VLA constructs task memory once at episode initialization and keeps it fixed throughout execution.
Abstract (from arXiv)
Vision-language-action (VLA) policies commonly place a multi-billion-parameter model inside the feedback loop, repeatedly recomputing task context whenever they predict a new action chunk. Yet task understanding changes far more slowly than the robot state. We introduce Fast-VLA, which invokes a frozen 4B VLM once per episode to contextualize the instruction and initial views as fixed, multi-layer key-value task memory. During execution, a lightweight online actor fuses learned time- and arm-indexed queries with current visual and proprioceptive feedback to form action queries. These queries retrieve task context from the fixed memory and predict action chunks without re-running the foundation model. Fast-VLA achieves 92.46% and 92.40% success on RoboTwin 2.0 Clean and Randomized, and 95.95% and 79.11% on LIBERO and LIBERO-Plus evaluation. Against three continually conditioned baselines on the same hardware, it reduces RoboTwin amortized policy inference time by 10.9-17.1x and requires 1.22 GiB of post-prefill online resident memory. Task-memory swaps further show that the actor relies on episode-matched task memory. These results show that task context can be established once while closed-loop control remains responsive throughout execution. Project page: https://garfield-kh.github.io/FastVLA
Related papers
- MIKASA-Robo-VLA: Benchmarking Memory in VLA Models for Long-Horizon Manipulation
- PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models
- Plan First, Act Later: Streaming Reasoning for Executable Memory in Robot Control
- EvoMem-VLA: State-Evolution Memory for Long-Horizon Robot Manipulation
- Recompose and Refine Latent Reasoning Flows for Vision-Language-Action Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-15 · Indexed by Pub-AI: 2026-10-16 UTC
Fast-VLA: You Only Think Once for Robot Manipulation
Preprint posted: 2026-10-15 · Original record ↗
- arxiv
- 2610.21389
Explicit source versions (1)
- v1 · 2026-10-15
- Preprint available
Source date: 2026-10-15 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Vision-language-action (VLA) policies commonly place a multi-billion-parameter model inside the feedback loop, repeatedly recomputing task context whenever they predict a new action chunk. Yet task understanding changes far more slowly than the robot state. We introduce Fast-VLA, which invokes a frozen 4B VLM once per episode to contextualize the instruction and initial views as fixed, multi-layer key-value task memory. During execution, a lightweight online actor fuses learned time- and arm-indexed queries with current visual and proprioceptive feedback to form action queries. These queries retrieve task context from the fixed memory and predict action chunks without re-running the foundation model. Fast-VLA achieves 92.46% and 92.40% success on RoboTwin 2.0 Clean and Randomized, and 95.95% and 79.11% on LIBERO and LIBERO-Plus evaluation. Against three continually conditioned baselines on the same hardware, it reduces RoboTwin amortized policy inference time by 10.9-17.1x and requires 1.22 GiB of post-prefill online resident memory. Task-memory swaps further show that the actor relies on episode-matched task memory. These results show that task context can be established once while closed-loop control remains responsive throughout execution. Project page: https://garfield-kh.github.io/FastVLA