React When You Need To: Event-Triggered Asynchronous Inference for VLA Policies

Yansong Wu; Huaqing Li; Tianding Hou; Lingyun Chen; Alois Knoll

cs.ROSep 18, 2026

React When You Need To: Event-Triggered Asynchronous Inference for VLA Policies

Authors: Yansong Wu, Huaqing Li, Tianding Hou, Lingyun Chen, Alois Knoll

Organizations: Technical University of Munich, Germany · Mohamed Bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE

Abstract

Vision-Language-Action (VLA) models commonly predict action chunks, limiting their ability to react to environmental changes during execution. Existing asynchronous inference methods improve reactivity but typically rely on a fixed inference gap. In this paper, we propose an event-guided dynamic inference strategy that adapts the inference gap according to scene changes observed since the previous inference. Thereby, it simultaneously preserves motion consistency and prompt reactivity. Across static and dynamic real-world settings, our method consistently performs best, averaging 95% success and exceeding the strongest baseline by 55 percentage points. The code will be made publicly available upon acceptance. The project page is available at https://react-when-you-need-to.github.io/.

Figures & tables

Fig. 1 : Representative event-score distributions under static and dynamic conditions for Task 1 and Task 2 detailed in Sec. V-A2 . The dashed blue line marks the task-specific 90th percentile of the static distribution, used as the event threshold

\theta_{\mathrm{P90}}

Panel of Figure 2 — Fig. 2 : Qualitative comparison of synchronous execution and our method under repeated obstructions caused by a moving human hand during a pick-and-place task.

Setting		Value
VLA fine-tune steps		$30{,}000$
VLA fine-tune batch size		$32$
Action-chunk horizon $H$		$50$
Blending weight decay $\beta$		$1.2$
Demonstration sampling rate		$30\,\mathrm{Hz}$
Execution frequency		$30\,\mathrm{Hz}$

TABLE I : Configuration Details

Fig. 3 : Representative Task 2 execution under target-motion perturbation, showing the action-chunk update at $t=6.4\,\mathrm{s}$ and the complete TCP trajectory at task completion ( $t=8.2\,\mathrm{s}$ ).

TABLE II : Stage-wise and full-task success rates for Task 1.

Fig. 4 : Execution horizon during Task 1. Dashed lines mark task events. The shaded region indicates repeated obstruction-avoidance cycles.

Fig. 5 : Success rates for Task 2 ( Place onto a Moving Target ) across baseline methods, fixed-gap ablations, and our method.

Fig. 6 : Execution horizon during Task 2 under target perturbation. The shaded region indicates the plate in motion; dashed lines mark motion reversal and the subsequent chunk update.

Fig. 7 : Aggregate full-task success across three evaluation settings, with each stacked segment representing one setting.

Vision-Language-Action Models

Vision-language-action models turn visual observations and natural-language instructions into actions for robots and other embodied agents.

See all 1,194 papers

Understanding Asynchronous Inference Methods for Vision-Language-Action Models
May 4, 2026Ayoub AgouzoulRobotic Control Efficient VLM Inference
École polytechnique · Systems Group, ETH Zürich
How Fast Can I Run My VLA? Demystifying VLA Inference Performance with VLA-Perf
Feb 20, 2026Wenqi Jiang, Jason Clemons, Karu Sankaralingam +1Efficient VLA Model Inference On-Device Inference
NVIDIA Research
Dynamic Execution Commitment of Vision-Language-Action Models
May 12, 2026Feng Chen, Xianghui Wang, Yuxuan Chen +4VLM Robustness Efficient VLM Inference
University of Adelaide · Sichuan University · Shanghai Jiao Tong University +3

May 4, 2026J/K move · Enter open · S save

Understanding Asynchronous Inference Methods for Vision-Language-Action Models

Ayoub Agouzoul

École polytechnique · Systems Group, ETH Zürich

Robotic Control Efficient VLM Inference Efficient VLA Model Inference Robotic Manipulation Vision-Language-Action Models

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models offer a promising path to generalist robot control, but their inference latency causes observation staleness when generated actions are executed asynchronously. Several methods have been proposed concurrently to mitigate this problem: inference-time inpainting (IT-RTC), training-time delay simulation (TT-RTC), future-state-aware conditioning (VLASH), and lightweight residual correction (A2C2). Each takes a fundamentally different approach, but they have so far been evaluated independently with different codebases, base policies, and protocols. We present a systematic comparison of these four methods under controlled conditions. We develop two unified codebases that integrate all methods with harmonized library and dataset versions, and we benchmark them on the Kinetix suite with MLPMixer policies and on the LIBERO manipulation benchmark with SmolVLA, sweeping inference delays up to

d=20

control steps. A2C2's per-step residual correction is the most effective method on Kinetix, holding above 90% solve rate up to

d=8

, and also leads on LIBERO from

d=4

onwards. IT-RTC is competitive at low delays but degrades sharply under long chunks (

H=30

) and high delays. TT-RTC is the most robust training-based method: stable across

d_\max

choices, generalizes beyond its training delay distribution, and adds zero inference overhead. VLASH exhibits a clear low-delay vs. high-delay trade-off governed by the fine-tuning delay range

[0,d_\max]

. Code is available at https://github.com/TheAyos/async-vla-inference

Abstract

Vision-Language-Action Models

Explore similar work

Understanding Asynchronous Inference Methods for Vision-Language-Action Models

How Fast Can I Run My VLA? Demystifying VLA Inference Performance with VLA-Perf

Dynamic Execution Commitment of Vision-Language-Action Models