FlashVLA, a method for robot-control AI, reported faster action decoding while maintaining or improving task success in selected simulated and physical manipulation tests. In the main asynchronous benchmark, the method paired a higher reported success rate with substantially less time per step than the comparison system.
The work is an arXiv preprint, version 1, dated 27 August 2026. Its main evaluation centered on π0.5 and the LIBERO and RoboTwin 2.0 simulated benchmarks, with additional tests on SmolVLA and LingBot-VLA. RoboTwin covered 50 tasks in clean and randomized settings.
A faster loop for asynchronous control
FlashVLA keeps action chunks moving through a staggered-noise streaming buffer and applies chunk-wise causal attention. In plain terms, it works through the chunks in sequence while retaining information from the stream. After a warm-up phase, it emits one executable chunk at each step. The preprint reports up to 20 times lower per-step action-decoding latency.
The tests focused on asynchronous control, in which the next inference is launched early rather than waiting for the current action sequence to finish. The delay is counted in low-level action steps, allowing the evaluation to examine whether speed and task quality hold as that timing gap changes.
The biggest gains came on longer tasks
The reported LIBERO asynchronous action-count averages were based on 2,000 episodes, with time defined as per-episode completion time. At a one-step asynchronous delay, average success was 97.8% with FlashVLA, compared with 96.9% for synchronous π0.5. Time per step fell from 53.8 milliseconds to 22.1 milliseconds, a reported 2.43-times speedup.
RoboTwin showed a smaller timing gap but the same direction in the reported matched-delay comparison. Across its 50 tasks, FlashVLA reached 90.6% success versus 86.0% for synchronous π0.5, while time per step declined from 47.4 to 43.8 milliseconds.
When the tested asynchronous delay ranged from one to four low-level action steps, LIBERO success stayed between 97.5% and 98.3%, while the reported speedup ranged from 2.43 to 2.62 times. At a four-step delay on RoboTwin, success was 89.8% for FlashVLA versus 86.0% for synchronous π0.5, with a 1.08-to-1.10-times speedup. An ablation that kept the buffer but removed causal attention lost the reported asynchronous gains.
The quality comparison was not limited to asynchronous runs. In synchronous evaluation, FlashVLA averaged 97.9% success on LIBERO versus 96.9% for π0.5. On RoboTwin, it reached 90.8% versus 86.1% in clean evaluation and 90.2% versus 85.8% when conditions were randomized.
The largest reported difference appeared after RoboTwin tasks were grouped by horizon. Compared with π0.5, FlashVLA's average difference was minus 0.7 percentage points on short-horizon tasks, plus 4.3 points on medium-horizon tasks and plus 36.6 points on long-horizon tasks. For the long-horizon group, clean success was 90.8% versus 54.2%, while randomized success was 88.4% versus 51.8%.
Speed held across hardware and models
Timing tests on matched hardware also favored the streaming method. With two views on an RTX 4090, reported latency fell from 45.8 to 26.7 milliseconds, and with three views it fell from 55.4 to 36.8 milliseconds. On an RTX 5090, the corresponding changes were 37.0 to 20.3 milliseconds and 44.8 to 27.1 milliseconds.
A separate responsiveness measure tracked time to first action and expected time to react. FlashVLA recorded 37.1 milliseconds for the first measure, compared with 80.0 for π0.5 and 62.1 for FASTER. Expected reaction time was 70.4 milliseconds, versus 130.0 and 112.1 milliseconds, respectively. The reported reductions against FASTER were 1.7 times and 1.6 times.
Tests on other policy architectures showed a similar pattern in the reported settings. SmolVLA latency fell from 19.7 to 10.1 milliseconds, a 1.95-times speedup, while LingBot-VLA latency fell from 70.6 to 25.1 milliseconds, a 2.81-times speedup. SmolVLA success was 80.1% synchronously and 79.5% with a one-step asynchronous delay. LingBot-VLA scored 88.6% synchronously and 89.3% asynchronously, compared with 85.2% for its baseline.
A physical-arm test adds a caveat
The method was also tried on a physical robot. The real-world evaluation covered three manipulation tasks, with 50 teleoperated demonstrations and 15 trials per task. Every asynchronous method used a two-step delay and the same system-level optimizations, making those details part of the reported setup.
Across the three tasks, FlashVLA's reported average score was 84.4%, compared with 80.0% for both synchronous π0.5 and RTC and 75.6% for naive asynchronous π0.5. The reported completion-time speedup averaged 1.3 times versus synchronous π0.5 and 1.2 times versus RTC. On RTX A4000 hardware, the paper reports 67.3 milliseconds of inference latency and 30-hertz control with the two-step delay.
The acknowledgment credits Professor Xiaolong Wang's lab with providing the Franka robot arm used in the real-world work.
What the numbers leave open
The physical timing results require care because completion time was averaged only over successful trials. Failed trials that ended early were therefore excluded from that comparison.
FlashVLA also has a one-time cold-start cost of the buffer length minus one warm-up step at the start of an episode. The document says that cost is amortized over multi-second rollouts but becomes more noticeable on very short tasks.
These findings are comparative engineering results within a defined test scope. They cover selected architectures, benchmarks, tasks, hardware configurations and runtime optimizations, so they do not establish performance beyond those settings. The reported comparisons also came without confidence intervals or other inferential uncertainty estimates.
Paper data and sources
Original title: FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference
Authors: Zekai Li, Jiaming Tang, Zhijian Liu
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-27
DOI: Not available
Original paper · Full text