UCLA's ChairNav brings long-horizon autonomous wheelchair navigation to real sidewalks
zhoubolei · x · 2026-10-30
Bolei Zhou's team at UCLA introduces ChairNav, exploring how vision-based navigation foundation models can enable long-horizon autonomous wheelchair navigation in complex real-world sidewalk environments.
The approach uses a three-stage pipeline: pre-training adapts delivery-robot data (aligning camera views and motion targets), mid-training learns wheelchair driving from human demonstrations, and post-training turns rider takeovers into preference supervision for personalization. The team built a dedicated wheelchair platform and ran real-world long-horizon evaluations across multiple routes, comparing autonomous vs. rider-control policies, including crossings and waiting. Paper: arXiv:2610.24012.
More from Embodied
- Foxconn's robotics play: in-house arms, humanoids and models, plus third-party robots — lukas_m_ziegler · 2026-10-30
- SplatXR: open-source Unity toolkit brings 3D Gaussian splats to Quest 3 and Vision Pro — Scobleizer · 2026-10-30
- Stanford & Northeastern unveil Retriever, a time-aware programming framework for long-horizon robots — yining_hong · 2026-10-30
- Mentra engineer on smart glasses in weak-network job sites: local STT offline, calls need just 1-1.5 Mbps — caydengineer · 2026-10-30
- Neuralink Prototype: Think "Hey Neuralink" to Trigger a Grok Computer-Use Agent — Baconbrix · 2026-10-30
- orcs v0.1.1 ships: toolkit for training privileged humanoid controllers with SONIC LoRA adaptation — kevin_zakka · 2026-10-30