Xiaoyang Xing and coauthors introduce APS-VLA for faster robotic control

Xiaoyang Xing and coauthors introduce APS-VLA for faster robotic control

Xiaoyang Xing and coauthors introduce APS-VLA, a way to reduce the visual information a vision-language-action (VLA) policy processes before generating robot actions. Its Decision Influence Estimator selects tokens using current visual observations, language instructions and robot proprioception, and is trained on robot demonstrations. The pretrained VLA backbone stays frozen; at deployment, the estimator selects and packs tokens in one step, and an auxiliary predictor used during training is removed. The authors report experiments with OpenVLA-OFT on LIBERO, and with π0.5 and XR-1 on RoboCasa365. Across five real-world tasks on two dual-arm platforms, they report mean policy inference latency falling from 137.25 to 93.25 ms, a 1.47× speedup, while average success was 61%, compared with 60% for the unpruned baseline. Escape Meridian interpretation: in these evaluations, the reported speedup came with a one-percentage-point difference in aggregate success, while the policy parameters and native action-generation procedure stayed unchanged. The abstract reports results from these evaluations, not a guarantee of similar performance elsewhere. It gives aggregate success rates, not the rate for each real-world task.

Publication: 0a1655cf-a502-4bd4-8c7f-4c0812b94519

attribution — pass: The reading attributes APS-VLA and its reported results to Xiaoyang Xing and coauthors. This matches the arXiv author list, and the results stay framed as author reports rather than established fact.

Authors:Xiaoyang Xing, Chun-Mei Feng, Anthony Cohn, Xiaoliang Gong, Asoke Nandi

summary — uncertain: The summary names the benchmark experiments (OpenVLA-OFT on LIBERO; π0.5 and XR-1 on RoboCasa365) but leaves out their reported outcome: acceleration with modest changes in task success. Only the favorable real-world result (61% vs 60%) is quantified, so readers may infer success was unaffected everywhere. Smallest repair: add that the authors report inference acceleration with modest changes in task success on these benchmarks. The reading also never says plainly why lower latency matters for robot control. Terms such as token, proprioception and inference latency should get term-library tooltips.

demonstrate inference acceleration with modest changes in task success

date — pass: originalPublished is 2026-11-09. This matches the arXiv v1 submission date, not the later feed date of 2026-11-11.

[Submitted on 9 Nov 2026]

support — pass: The claims, latency figures, speedup and success rates match the abstract. The frozen backbone, the removed auxiliary predictor and the unchanged policy parameters are all stated in the source. The Escape Meridian interpretation is labeled and limited to the reported evaluations. The question about per-task success follows from the abstract giving only averages, without implying an answer.

APS-VLA reduces mean policy inference latency from 137.25 to 93.25 ms, achieving a $1.47\times$ speedup with an average success rate of 61%, compared with 60% for the unpruned baseline

media — pass: The selected media is the verified direct link to the original arXiv abstract page. It matches the source URL and is the most direct way to reach the paper.