FlowPilot: New AI navigates sidewalks 40% safer with human feedback

Researchers from the robotics community have published a paper introducing FlowPilot, a novel navigation policy designed for autonomous long-horizon sidewalk navigation using only a single monocular RGB camera. This is a significant step for real-world micro-mobility applications like robotic food delivery and assistive wheelchairs, where lightweight perception is critical. FlowPilot leverages anchored flow matching—a generative action representation—to pre-train on large-scale fleet data, capturing the multimodal distribution of sidewalk behaviors. This addresses the compounding errors typical of imitation learning.

To bridge the gap between imitation and true social alignment, the team added a human-in-the-loop preference learning scheme. By collecting a small amount of human intervention data, the model learns counterfactual reasoning and social compliance, such as yielding to pedestrians or navigating unpredictable terrain. In simulation, FlowPilot achieved a 42% success rate and 66% route completion. Real-world tests showed even greater gains: the human-preference tuned version (FlowPilot-HP) cut the intervention rate by 40.0% and near-intervention rate by 52.1% relative to the base model, demonstrating robust performance in diverse sidewalk environments.