FlowPilot

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

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

The training architecture is available as the model=flowpilot recipe in VisNavKit — FastViT-MA36, a 4-layer scene transformer, 64 k-means anchors and the anchored rectified-flow decoder with Beta(1.5, 1) time sampling, composed from swappable Hydra groups, with ONNX export and an open-loop benchmark:

uv run visnavkit-train dataset=torch model=flowpilot

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