Research Analyzer
← Back ICRA 2026

ABPolicy: Asynchronous B‑Spline Flow Policy for Real‑Time and Smooth Robotic Manipulation

Fan Yang, Peiguang Jing, Kaihua Qu, Ningyuan Zhao, Yuting Su

PDF

AI summary

Key figure (auto-extracted from paper)
ABPolicy achieves smoother, more responsive robotic manipulation by combining asynchronous inference with B-spline control points and continuity-constrained refitting.
Asynchronous inference B-spline parameterization Flow matching Smooth manipulation Real-time control Bidirectional prediction

Problem

Synchronous inference in raw action spaces causes intra-chunk jitter, inter-chunk discontinuities, and execution stalls, which degrade smoothness and responsiveness to dynamic environments.

Approach

The method predicts continuous B-spline control points using a flow-matching model and enforces trajectory continuity through bidirectional prediction and local refitting optimization, all while running inference asynchronously to hide latency.

Key results

  • Inherent intra-chunk smoothness via continuous B-spline control points
  • Seamless inter-chunk continuity through bidirectional prediction and refitting
  • Real-time responsiveness via asynchronous inference that eliminates execution stalls
  • Higher success rates and reduced trajectory jerk across seven static and dynamic manipulation tasks

Why it matters

Provides a practical, latency-free control framework for robots operating in dynamic, real-world environments where smoothness and reactivity are critical.

Abstract

Robotic manipulation requires policies that are smooth and responsive to evolving observations. However, syn- chronous inference in the raw action space introduces several challenges, including intra-chunk jitter, inter-chunk disconti- nuities, and stop-and-go execution. These issues undermine a policy’s smoothness and its responsiveness to environmental changes. We propose ABPolicy, an asynchronous flow-matching policy that operates in a B-spline control-point action space. First, the B-spline representation ensures intra-chunk smooth- ness. Second, we introduce bidirectional action prediction cou- pled with refitting optimization to enforce inter-chunk conti- nuity. Finally, by leveraging asynchronous inference, ABPolicy delivers real-time, continuous updates. We evaluate ABPolicy across seven tasks encompassing both static settings and dy- namic settings with moving objects. Empirical results indicate that ABPolicy reduces trajectory jerk, leading to smoother motion and improved performance. Project website: https: //teee000.github.io/ABPolicy/.

Index terms

Imitation Learning Deep Learning in Grasping and Manipulation Learning from Demonstration

Related papers