From Imitation to Alignment: Human-Preference Flow Policies for Long-Horizon Sidewalk Navigation

2026-06-10 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

FlowPilot is a novel mapless navigation policy designed for robust, efficient long-horizon sidewalk navigation using only a monocular RGB camera. Developed by UCLA, this framework addresses limitations of traditional imitation learning, such as compounding errors and poor social compliance. It employs anchored flow matching for policy pre-training on large-scale robot fleet data, capturing diverse, multimodal navigation behaviors. A subsequent human-in-the-loop preference learning scheme fine-tunes the policy with minimal human intervention data, enhancing counterfactual reasoning and social compliance. In simulation, FlowPilot-Base achieved a 42% success rate and 66% route completion. Real-world experiments demonstrated FlowPilot-HP, with preference fine-tuning, further improved robustness and social compliance, reducing the Intervention Rate by 40.0% and Normalized Intervention Rate by 52.1% compared to the base model.

Key takeaway

For Robotics Engineers deploying micro-mobility robots in complex sidewalk environments, relying solely on imitation learning for navigation policies risks poor social compliance and compounding errors. You should integrate human-in-the-loop preference learning to fine-tune pre-trained models, ensuring robot-specific precision and adherence to social norms. This approach, demonstrated by FlowPilot's 40.0% IR reduction, will significantly improve real-world robustness and reduce necessary human interventions.

Key insights

Bridging imitation learning with human preference alignment significantly enhances robot navigation robustness and social compliance.

Principles

• Anchored flow matching models multimodal action distributions.
• Preference learning refines policies for robot-specific precision.
• Gated cross-attention improves context utilization, mitigating goal-driven shortcuts.

Method

Pre-train a policy using anchored flow matching on large datasets, then fine-tune it with reward-free preference learning from human interventions, regularizing towards the initial policy.

In practice

• Apply anchored flow matching for multimodal trajectory prediction.
• Use human-in-the-loop preference learning for robot-specific adaptation.
• Integrate gated attention to enhance scene context understanding.

Topics

• Sidewalk Navigation
• Autonomous Robotics
• Imitation Learning
• Preference Learning
• Flow Matching
• Monocular Vision

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• From Imitation to Alignment: Human-Preference Flow Policies for Long-Horizon Sidewalk Navigation — FlowPilot combines anchored flow matching with human-in-the-loop preference learning for robust, socially compliant sidewalk navigation.
• A Mechanistic Analysis of Sim-and-Real Co-Training in Generative Robot Policies — Co-training effectiveness in robotics stems from structured representation alignment and importance reweighting.
• Multi-Agent Reinforcement Learning for Safe Autonomous Driving Under Pedestrian Behavioral Uncertainty — Co-training SDCs with personality-driven MARL pedestrians significantly improves safety by exposing vehicles to realistic, uncertain human behaviors.
• Beyond Waypoints: A Trajectory-Centric Waypointing Paradigm for Vision-Language Navigation — Grounding waypoints in executable trajectories resolves planning-control inconsistencies and improves reachability in Vision-Language Navigation.
• X-Diffusion: Training Diffusion Policies on Cross-Embodiment Human Demonstrations — X-Diffusion selectively integrates noisy human demonstrations into robot policy training to overcome embodiment mismatches.
• PLAN-S: Bridging Planning with Latent Style Dynamics for Autonomous Driving World Models — PLAN-S enhances autonomous driving LWMs by explicitly modeling style-conditioned spatial costs for improved controllability and safety.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.