MapFly-Agent

Reference policy of the MapFly benchmark: Qwen3-VL-4B with starVLA's QwenOFT action head, plus a stop and progress head. Code is starVLA/examples/uav/ in that repository.

Checkpoint

P1-R0 on OSM maps, 80k steps (main rows of Tables I–III):

Folder	Head	Test Seen SR	Test Unseen SR
mapfly_agent_oft_p1r0/	QwenOFT	86.6	62.7

mapfly_agent_oft_p1r0/
  config.yaml
  dataset_statistics.json
  checkpoints/steps_80000_pytorch_model.pt

Download into $PLAYGROUND/Checkpoints/ (default starVLA/playground/Checkpoints):

huggingface-cli download EzGuYan/MapFly-Agent mapfly_agent_oft_p1r0 --local-dir playground/Checkpoints/mapfly_agent_oft_p1r0

Serve and evaluate from the MapFly checkout:

CKPT=$PLAYGROUND/Checkpoints/mapfly_agent_oft_p1r0/checkpoints/steps_80000_pytorch_model.pt \
  bash starVLA/examples/uav/eval_files/run_policy_server.sh
ROLE=unseen TAG=oft_p1r0 bash starVLA/examples/uav/eval_files/run_eval_split.sh

The other paper rows (QwenGR00T, other tracks) are trained with HEAD / VARIANT in starVLA/examples/uav/README.md; those checkpoints are not released here.

config.yaml names the base VLM as Qwen/Qwen3-VL-4B-Instruct. Put a copy under playground/Pretrained_models/ if your serving setup still expects a local tree.

License

MIT. The Qwen3-VL weights inside the checkpoint follow Qwen's license.