Understanding Asynchronous Inference Methods for Vision-Language-Action Models

Summary

This paper finds that A2C2 is the strongest high-delay method for asynchronous VLA control, while TT-RTC has the lowest runtime cost when it works. The study is mainly a controlled benchmark of existing methods, with two unified codebases and matched evaluation settings.

Problem

VLA policies often predict action chunks, but inference takes enough time that the robot acts on an old observation by the time a chunk is ready.
Synchronous execution avoids stale observations by pausing the robot, which lowers the control rate and can hurt fast manipulation or dynamic control.
Prior methods for this delay problem were tested in separate codebases with different base policies, datasets, and protocols, so their trade-offs were hard to compare.

Approach

The paper compares four methods under shared implementations: IT-RTC, TT-RTC, VLASH, and A2C2.
IT-RTC keeps the already-committed action prefix fixed and inpaints the rest of the action chunk during flow-matching inference.
TT-RTC trains the policy with simulated delays, so it learns to condition on an action prefix and predict the remaining actions with no added inference cost.
VLASH estimates the robot state at execution time by rolling the state forward, then conditions the policy on that future state.
A2C2 runs a small correction model at every control step and adds a residual action to the stale base-policy action.

Results

On Kinetix with 10 environments and chunk size H=16, A2C2 stayed above 90% solve rate up to delay d=8, while the naive asynchronous baseline fell below 40%.
On Kinetix with H=30 and delays up to d=15, A2C2 led overall; TT-RTC generalized well across d_max values of 4, 8, and 15; IT-RTC gave little gain over naive at high delay.
On LIBERO with SmolVLA, 40 tasks, H=50, and delays d∈{0,1,2,4,8,15,20}, A2C2 led from d≥4 and reached about 58% success at d=20.
At LIBERO d=20, the naive baseline was about 10-12%, IT-RTC about 20%, TT-RTC about 25% for d_max=8 and about 33% for d_max=4, while VLASH with d_max=8 or 16 stayed near 55-56%.
LIBERO VLASH used ground-truth future states because the benchmark state and action dimensions do not match, so its high-delay results are an upper bound for a deployable version.
On an RTX 3090 for LIBERO, naive SmolVLA inference took 405.2 ms per chunk, TT-RTC took 402.7 ms, IT-RTC took 469.7 ms, and A2C2 serial inference took 412.4 ms; the A2C2 residual head alone took 7.27 ms.

Link

https://arxiv.org/abs/2605.08168v1