UCLA's FlowPilot uses human preferences to cut sidewalk robot interventions by 40%
zhoubolei · x · 2026-11-06
UCLA's Vision and Autonomy Intelligence Lab (Bolei Zhou's team) presents FlowPilot at CoRL 2026: a mapless, monocular-camera policy for long-horizon sidewalk navigation, validated on Coco Robotics hardware.
Method
- Anchored Flow Matching: a conditional flow-matching policy anchored to clustered prototypical behaviors, with gated cross-attention grounding decisions in scene context and suppressing goal-driven shortcuts.
- Reward-free preference alignment: human-in-the-loop fine-tuning from corrective interventions that aligns the policy with socially compliant behavior while preserving imitation priors.
Results
- FlowPilot-Base hits 42% success rate and 66% route completion in simulation.
- The preference-tuned FlowPilot-HP cuts real-world intervention rate by 40.0% and normalized intervention rate by 52.1%.
Related event: UCLA's FlowPilot Cuts Sidewalk Robot Intervention Rate by 40%(2 posts)→
More from Embodied
- YC backed 25 robotics startups last batch as VCs call robotics the new SaaS moment — Scobleizer · 2026-11-06
- Apple Vision Pro Positano scene hits 65fps on M2 with boats, yachts and 40 seagulls — Scobleizer · 2026-11-06
- Dev builds a budget-friendly, somewhat buggy DIY Vision Pro — aaraalto · 2026-11-06
- Developer builds a computer hat with hand-gesture recognition for screen-free control — aaraalto · 2026-11-06
- How realistic must simulated failures be? Low fidelity suffices with little data — mangahomanga · 2026-11-06
- Johns Hopkins study compares four failure data sources for robot reward models — mangahomanga · 2026-11-06