Horizon Robotics' VITAL hits 90.8 PDMS with a single denoising step at 3.29x inference speed

HorizonRobotics · hf · 2026-10-15

• Horizon Robotics, with Shanghai Innovation Institute, Fudan, NUS and Zhejiang University, released VITAL (Vision-Induced Tokens for Action Learning), a world-action model framework for end-to-end driving.
• Key idea: instead of training action experts on trajectories defined independently of visual dynamics, VITAL derives action targets directly from visual motion — a visually derived motion scaffold plus a compact residual that encodes the centimeter-scale trajectory error left by that scaffold.
• A flow-matching action expert jointly predicts both components from the current image, ego state, and navigation command, without generating future images at inference; the residual coordinates also enable bounded RL post-training that refines execution while keeping the scaffold fixed.
• Results: a single denoising step scores 90.8 PDMS on NAVSIM v1 navtest (within 0.13 of ten steps, at 3.29x end-to-end inference speed); without retraining it reaches 91.1 EPDMS on NAVSIM v2 and 38.6 on navhard.

Original post →

More from Embodied

• SimDist at RSS2026 lifts failing robot policies to ~90% success with 15-30 minutes of data — KyleMorgenstein · 2026-10-15
• AnyRank: Sequential Testing Ranks Robot Policies Reliably With Up to 50% Fewer Trials — ZeYanjie · 2026-10-15
• Natura-1 Robot Model Demo: Zero-Shot Generalization With Autonomous Retries in One Take — ZeYanjie · 2026-10-15
• SafeDriveVLA: world model 'dreaming' guards driving VLAs against unsafe navigation commands — UCI6055 · 2026-10-15
• Grain-levelling robot keeps workers out of ship holds, cutting engulfment and fall risks — LexiLove · 2026-10-15
• Pi-0.5 can post-train to grab a pen, but legs-balance coordination is the hard part — Darpinian · 2026-10-15

Embodied channel →