FlowPilot

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

Conference on Robot Learning (CoRL) 2026

_{GPS-guided long-horizon navigation · sidewalk lane keeping · obstacle avoidance · pedestrian awareness · night driving}

_{Closed-loop rollouts in SidewalkBench-GS · Gaussian-splat reconstructions of real sidewalks}

Training recipe

The FlowPilot architecture ships as the model=flowpilot recipe in VisNavKit: a FastViT-MA36 encoder, a 4-layer scene transformer, 64 k-means anchors and an anchored rectified-flow decoder with Beta(1.5, 1) time sampling. The components are swappable Hydra groups, and the kit includes ONNX export and an open-loop benchmark.

uv run visnavkit-train dataset=torch model=flowpilot

Release

Code, checkpoints and the human-preference alignment stage will be released here.