LCVN: language-conditioned visual navigation with 39k trajectories, two world-model approaches compared
机器之心 · wechat · 2026-11-08
Researchers from University of Washington and others introduce LCVN, a language-conditioned visual navigation task: given only a starting image and a text instruction, models must predict the full action sequence offline, with no real environment feedback for error correction. The dataset has 39,016 trajectories and 117,048 human-verified instructions across five sources (GoStanford, ReCon, SCAND, HuRoN, TartanDrive), with three instruction styles per trajectory, scoring 5.2/6.0 in human evaluation. Two approaches are compared: LCVN-WM+LCVN-AC trains a policy inside a world model's imagination (43% in-domain success, best visual coherence with SSIM@8 of 0.293), while LCVN-Uni, built on Anole-7B, jointly predicts actions and frames as unified tokens and generalizes better to unseen off-road terrain (27% vs 19%). Removing language conditioning drops success from 43% to 35%. A WorldNav track is live at RoboWorld Challenge 2026.
More from Embodied
- Galaxea Co-founder Hang Zhao's CoRL Keynote: Split Robot Intelligence into System 2/1/0 by Timescale — YuXiang_IRVL · 2026-11-13
- Apparent OpenAI hardware unboxing photo leaks: robot arm plus handheld console — yungcontent · 2026-11-13
- FINGR dexterous hand solves a 2x2 Rubik's Cube with continuous finger tricks, wins Best Paper at CoRL 2026 workshop — Scobleizer · 2026-11-13
- Fei-Fei Li amplifies new Atlas demo: 'It is very hard to explain how cool Atlas is' — drfeifei · 2026-11-13
- GXO to deploy up to 1,000 Apptronik Apollo humanoid robots in warehouses over four years — CyberRobooo · 2026-11-13
- Waymo robotaxi spotted quietly obeying a 'no right turn on red' sign humans often miss — reed · 2026-11-13