Freebar improves real-time asynchronous video line action success rates
Barrier-Constrained Guidance For Real-time Asynchronous VLA Execution
Robotics
Summary
Real-time systems that predict actions in chunks often face problems with mismatched boundaries between what the system predicts and what actually happens. The authors introduce FreeBar, a simple add-on that fixes this issue without needing extra training or complex model changes. It uses a special consistency check inspired by control theory to align predicted actions with executed ones smoothly. This method reduces errors and speeds up decision-making in dynamic environments like robotics.
What this means in practice
- •For robotics engineers: Improve robot motion planning by reducing delays and alignment errors in asynchronous action prediction.
- •For real-time ai developers: Enhance real-time AI systems that generate action sequences by maintaining alignment between predictions and execution with low latency.
Authors
Wenbo Yu, Jifan Li, Wenke Xia, Yuchen Li, Yizhuo Zhang, Pei Ren, Zhengping Che, Di hu
Abstract
While asynchronous execution enables VLA models to operate in real-time environments by predicting the next chunk during execution, it inevitably introduces boundary misalignment with the committed action stream. Existing delay-aware solutions address this at the prohibitive cost of extensive retraining, architectural modifications, or fixed-prefix inpainting that inflates computation when low latency is vital. To address these problems, we propose FreeBar, a training-free real-time inference plugin that enables robust asynchronous execution with only marginal additional inference cost by casting chunk alignment in flow-based VLAs as barrier-guided action-space constraint. Specifically, FreeBar introduces a lightweight Control Barrier Function (CBF)-inspired consistency constraint between consecutive action chunks and solves the resulting correction in closed form, avoiding costly backpropagation through the VLA action head. By progressively tightening the consistency bound during generation, FreeBar aligns generated chunks with the committed execution stream while preserving early-stage generative and within-chunk future action flexibility,. Across Kinetix, RoboCasa, and diverse real-world robotic platforms with multiple VLA backbones, FreeBar improves success rates over naive asynchronous execution by 10%--15% on average, with larger gains in highly dynamic scenarios. It also preserves real-time computational efficiency, decreasing inference latency per action chunk over 45% than RTC. Additionally, as a universal plugin, FreeBar remains robust across distinct latency environments, base policies, and hyperparameter settings.