Research map / Vision-language and multimodal / Robot action and navigation

FASTER: Rethinking Real-Time Flow VLAs

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FASTER, a horizon-aware flow-sampling schedule plus streaming client-server interface that makes the first action of flow-based VLAs available after a single denoising step, cutting reaction latency on real robots.

Problem

Asynchronous real-time inference for flow-based VLAs optimizes trajectory smoothness but neglects reaction latency to environmental changes, because a constant denoising schedule forces all sampling steps to finish before any action is dispatched.

Approach

A Horizon-Aware Schedule assigns index-dependent flow timesteps so near-term actions reach their clean state within one sampling step, combined with streaming action dispatch and early stopping; trained by mixing this schedule with the constant schedule during fine-tuning, with no architecture change.

Key results

• On RTX 4060, FASTER reduces TTFA for X-VLA from 399.5 ms to 129.2 ms (3.09x speedup) and pi0.5 from 303.3 ms to 238.6 ms.
• On RTX 4090 with X-VLA, expected reaction time drops from 180.4 ms (async) to 78.1 ms (2.31x); FASTER dominates async deterministically (probability 1.00).
• Real-world table tennis: Sync fails to respond to incoming balls; FASTER achieves faster reactions and better scores than naive async and Training-time RTC on both GPUs.

Surprising findings

• Expected reaction time under asynchronous inference is only reduced by 0.5 times the inference latency versus synchronous, so upgrading to async yields limited reaction gains.

Limitations

• Potentially slight loss in action prediction accuracy due to accelerated sampling of later actions
• Evaluated mainly on a single low-cost robot platform with few real-world trials (10-35 rollouts per task)

Experimental setup

Datasets: AgileX Cobot Magic teleoperation data (table tennis, pick beverage, fold towel) LIBERO CALVIN ABC openvla/modified_libero_rlds

Benchmarks: LIBERO CALVIN

Metrics: TTFA minimum execution horizon s_min expected reaction time probability of faster reaction task completion score completion duration

Baselines: Sync inference Naive Async Training-time RTC

Builds on: real-time chunking (RTC) training-time RTC diffusion forcing flow matching

Closest papers by meaning

• Real-Time Execution of Action Chunking Flow Policies (NeurIPS 2025 · similarity 0.90)
• PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking (NeurIPS 2026 · similarity 0.90)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
• VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.89)
• ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.88)
• Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.88)