In simulated RoboTwin 2.0 tests, authors report that Fast-VLA ran inference 11.2× faster per environment step than the DM0.5 baseline, while its mean task-success rate was 1.02 percentage points below DM0.5’s 93.45%. The arXiv preprint builds task memory once at the start of an episode, then reuses it as a smaller actor adjusts movements from fresh camera and robot-state data.
One large model call prepares a smaller control loop
A vision-language-action (VLA) policy turns camera views and a language instruction into robot movements. Many VLAs run a large vision-language model (VLM) again for each new chunk of actions. Fast-VLA instead uses a 4-billion-parameter VLM once at the start of each episode. It encodes the instruction and initial camera views into a stored representation of the task and scene that the smaller actor can reuse throughout the episode.
A separate online actor uses current camera images and the robot’s joint-state feedback to read that memory and predict the next chunk of actions. It updates its control from new observations, while the large VLM does not reprocess the task for every chunk. In the RoboTwin tests, the online actor has 272 million parameters; the separate 4B VLM is used only to build the initial task-and-scene memory.
Benchmarks show speed gains and mixed success rates
RoboTwin 2.0 simulates two-arm robot manipulation tasks. The authors evaluated 100 episodes for each of 50 tasks in the Clean and Randomized settings, 10,000 rollouts total. They report 92.46% success in Clean and 92.40% in Randomized, a 92.43% mean. On the same hardware, with both systems predicting and executing 50 actions between replans, Fast-VLA averaged 1.437 milliseconds per environment step, including its initial VLM call; DM0.5 averaged 16.069 milliseconds. Fast-VLA’s mean success rate was 92.43%, compared with DM0.5’s 93.45%.
In the same RoboTwin comparison, the paper reports that Fast-VLA was 10.9× faster than Faster-WAM (HUST) and 0.09 percentage points lower in mean success. Against Fast-WAM, it was 17.1× faster and 0.60 points higher in mean success.
On separate LIBERO evaluations, the paper reports 95.95% success on standard LIBERO and 79.11% on LIBERO-Plus. On LIBERO-Plus, Fast-VLA trails pi0.5 and DM0.5 by 6.59 and 6.79 percentage points, respectively. These are benchmark results; the paper reports no physical-robot deployment result.
Mid-task instruction changes remain unaddressed
The authors say their method does not yet address a new instruction or changed goal arriving mid-task. The online actor uses fresh visual and robot-state data to adjust movements, but the high-level task memory remains fixed throughout the episode.
Image credit: Lexington Medical, Inc., via Wikimedia Commons; CC0 1.0. Illustrative manufacturing photo, not the Fast-VLA system.
Sources
- Kehong Gong et al., “Fast-VLA: You Only Think Once for Robot Manipulation,” arXiv, Oct. 15, 2026
- “RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation”
- “LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models”
- Wikimedia Commons, “Lexington Medical, Inc. Manufacturing Robot Arm.jpg” (CC0 1.0)