RACE: 4x Longer Action Chunks for VLA Robots, 5x Less Idle Time
POSTECH · hf · 2026-10-06
RACE: Reliable Action-Chunk Extension for VLA Models
VLA models are unified policies for robotic manipulation, but expensive inference forces robots to pause between policy calls, producing stop-and-go execution. Longer action chunks reduce calls but make execution unreliable.
POSTECH researchers find action errors within long chunks concentrate at transitions between subskills, growing sharply with chunk length—so transition timing is key.
RACE predicts transition timing from an auxiliary one-step denoising pass and conditions action generation on it, reducing errors at subskill transitions and enabling reliable long-chunk execution.
Results:
- In simulation, RACE beats fine-tuning at equal chunk length; with 2x longer chunks it surpasses recent SOTA efficient VLAs in success rate, staying competitive at 4x;
- On a real robot, 4x longer chunks cut stop-and-go idle time by about 5x while achieving higher success rate than fine-tuning at the same length.
Code and real-robot demo: github.com/Seonghoon-Yu/RACE-VLA
More from Embodied
- VC: almost none of real robot deployment data flows back into training — carrycooldude · 2026-10-06
- RealtimeWAM: One-Step Asynchronous World Action Model Delivers 25x Speedup with <1% Accuracy Loss — NanyangTechnologicalUniversity · 2026-10-06
- ACG-Bench Probes Dual-Arm VLA Generalization; AE-VLA Lifts Success from 3% to 21.5% — Zaibin Zhang · 2026-10-06
- ReSteer open-sourced: fixing VLA policies that ignore mid-execution instruction switches — siddkaramcheti · 2026-10-06
- DIY robotic backboard uses computer vision to make every beginner shot go in — TinfoilTricorn · 2026-10-06
- Xitac tactile sensor uses magnetotactic gel over 3D hall array to capture finger forces — seanmcdonaldxyz · 2026-10-06