eco-drone-policies

Astral eco — drone navigation policies

Trained RL navigation policies for the astral-us/eco drone platform (quadcopter, rover, fixed-wing). Each policy takes a 56-dim forward-depth state and outputs a 4-dim [vx, vy, vz, yaw_rate] action; see eco/drone/training/contract.py / fw_contract.py for the exact state layout.

These are optional. The platform's core, benchmarked result — the L5 zero-intervention fleet controller (eco/drone/common/L5.md) — is classical potential-field control with no model weights at all. These .onnx files back a separate, secondary learned-policy path (LearnedPlanner in reactive_planner.py) used mainly for the Isaac Sim demo renders; in the team's own testing it underperforms the classical controller on the L5 benchmark.

Files

Usage

Download into eco/drone/models/ (the filenames already match what vehicle_class.py expects as each class's default policy_onnx):

pip install huggingface_hub
python -c "
from huggingface_hub import hf_hub_download
for f in ['policy_v26rnn_dr.onnx', 'policy_rover_v2.onnx', 'policy_rover_v2.onnx.data', 'policy_fw.onnx']:
    hf_hub_download(repo_id='astralhf/eco-drone-policies', filename=f, local_dir='eco/drone/models')
"

Or with the CLI:

huggingface-cli download astralhf/eco-drone-policies --local-dir eco/drone/models

Training your own

See eco/drone/training/justfile — train / train-ma / train-fw recipes (PPO from scratch or with BC warm-start, run on a CUDA GPU), and pull / pull-ma / pull-fw to fetch the result. The fixed-wing policy here is an early checkpoint; retraining for longer or with tighter domain randomization is expected to improve on it.