When Does Asynchronous Action Chunking Help? Latency, Precision and Stalls for VLAs on Edge CPUs

This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Vision-language-action (VLA) policies predict chunks of future actions, and asynchronous inference computes the next chunk while the robot executes the current one. On edge CPUs a single chunk can take seconds, comparable to or longer than the time the chunk lasts, and it is not obvious when overlapping inference with execution still helps. We give a closed-form model of how often the robot stalls under synchronous and asynchronous chunking, with chunks aligned to executed actions. It identifies three regimes: when inference takes at most half a chunk, a just-in-time request never stalls; between half a chunk and a full chunk, asynchronous execution stalls less than synchronous; beyond a full chunk it stalls more. Measurements on a Raspberry Pi 5 match the model within one percentage point. We then test whether fewer stalls mean more completed tasks, on LIBERO-Spatial with SmolVLA and a simulated inference delay. They do not. At a delay of a quarter chunk, asynchronous execution keeps the success rate and finishes successful episodes 17 steps sooner. At 0.64 of a chunk it stalls less, as the model predicts, but succeeds in 4 of 30 episodes against 14 of 30 for synchronous execution (paired p = 0.006). A request rule should therefore be asynchronous only inside the stall-free regime and synchronous outside it. Weight precision decides which regime an edge CPU is in: on a Raspberry Pi 5, 8-bit weights cut a chunk of the SmolVLA base model from 11.9 s to 3.3 s. On a larger LIBERO-tuned SmolVLA checkpoint, evaluated on an Apple M4, 8-bit weights cause no measurable change in task success (28/50 and 26/50 episodes for two 8-bit configurations against 24/50 for the 32-bit reference; paired p >= 0.29). Code and data: https://github.com/SkidGod4444/sapient (release v0.6.3)