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) unveils FlowPilot, a mapless, monocular-camera navigation policy for long-horizon sidewalk robots, to be presented at CoRL'26.
- Imitation to alignment: pretrain on large-scale offline demos, then fine-tune with a small amount of human-preference data for safe, socially compliant behavior.
- Anchored Flow Matching: conditional flow matching anchored to clustered prototypical behaviors, with gated cross-attention to suppress goal-driven shortcuts.
- Reward-free human-in-the-loop preference learning from corrective interventions while preserving imitation priors.
- Results: FlowPilot-Base hits 42% success rate and 66% route completion in simulation; preference-tuned FlowPilot-HP cuts real-world intervention rate by 40.0% and normalized intervention rate by 52.1% (validated on Coco Robotics hardware/data).
Related event: UCLA's FlowPilot Cuts Sidewalk Robot Intervention Rate by 40%(2 posts)→
More from Embodied
- Developer Validates AME-2 on Humanoids: 70cm Foot-Only Climbing, 1.2m Gap Traversal in Sim — ChongZzZhang · 2026-11-06
- Wayve releases Learned Safety strategy after a decade of end-to-end AV research — alexgkendall · 2026-11-06
- Generalist AI demos GEN-1.5 with millimeter-scale O-ring dexterity tasks — ChongZzZhang · 2026-11-06
- Florida deploys robot bunnies to lure invasive pythons that wiped out 99% of raccoons — IsForAt · 2026-11-06
- Lamport Robotics' Pincer1: a Six-Legged Robot Wallet That Signs Only When You're in the Room — R4nsack · 2026-11-06
- RoboColosseum tests 5 VLA policies across 4 embodiments; environment diversity is the new bottleneck — DJiafei · 2026-11-06