NavGPT-3 Reaches Human-Level Navigation: 90.43 SR on RxR-CE vs Human 90.4
Metacognition-AI · hf · 2026-10-15
NavGPT-3 connects LLM reasoning with low-level control via an OS-like runtime where reasoning, acting, and monitoring run as threads with their own contexts, tools, and permissions, scheduled with interruption and thread switching. Its 8B NavGPT VLA (trained on 19.28M examples) hits 74.51 SR on R2R-CE and 78.19 on RxR-CE; the full system sets SOTA on R2R-CE (81.51 SR) and matches human followers on RxR-CE (90.43 vs 90.4 SR, 78.47 vs 77.7 nDTW) at 1m22s per episode vs 3min for humans. The VLA cuts minimum reaction time from 3–19s per LM decision to 0.5–1s per step. Models, code, and eval records will be released.
More from Embodied
- Scoble: Tesla FSD 14.3 update is spectacular even on 8-year-old Hardware 3 cars — Scobleizer · 2026-10-15
- Dev puts a Muse AI agent on the wall with Pi 5 and laser projector — alexandr_wang · 2026-10-15
- Figure 02 humanoid crosses obstacles blind, powered by an RL controller — jonstephens85 · 2026-10-15
- HapticWorld: A world simulator that lets teleoperators feel joint torque in real time — Scobleizer · 2026-10-15
- Rice's RoboTok Talk: Mining Internet Videos for Dexterous Robot Learning — ceciletamura · 2026-10-15
- SimDist at RSS2026 lifts failing robot policies to ~90% success with 15-30 minutes of data — KyleMorgenstein · 2026-10-15