B-spline Policy proposes faster, smoother robot manipulation

Starting on July 15, HaoyuXiong introduced B-spline Policy (BSP), a robot manipulation method centered on continuous action representation. The method drew attention because it targets a common weakness of visuomotor policies: even when tasks are completed successfully, robot motions can be overly slow and transitions between phases can feel unnatural. HaoyuXiong argued that evaluation should not focus only on success rate, but also on speed.

Core idea

According to HaoyuXiong, BSP does not predict future actions as fixed-frequency, discrete chunks. Instead, it parameterizes actions as a continuous B-spline curve. During execution, the robot can change the time step used to sample from that curve, which allows the same policy to run at different speeds across different stages of a task rather than being locked to one fixed action frequency.

Why it matters

In discussion shared by duyilun, a key reason visuomotor policies become slow is action chunking: the policy predicts waypoints at a fixed frequency and effectively treats all task phases similarly. BSP is presented as an attempt to address that mismatch with a continuous representation that better matches varying speed requirements. HaoyuXiong summarized the expected effect as making robotic arm motions faster, smoother, and more stable.

2026-07-15 ~ 2026-07-16 · 5 related posts

• [source] B-spline Makes Robot Arms Faster and Smoother — Haoyu_Xiong_ · 2026-07-15
• B-spline Policy Enables Variable-Speed Robot Motion — du_yilun · 2026-07-16
• B-Spline Policy Speeds Up Robots — chris_j_paxton · 2026-07-16

2 near-duplicate retellings: Haoyu_Xiong_ · Haoyu_Xiong_