AirVLA FT-DAG (pi0, step 17500, on-policy corrective)

DAgger-style on-policy corrective fine-tune: expert completions recorded from states the policy itself reaches. Improves terminal precision but does not unlock the grasp, and significantly degrades target selection.

Part of the AirVLA MSc dissertation campaign: adapting a pretrained pi0 vision-language-action policy to a quadcopter with a 2-DoF arm and parallel gripper, evaluated in MuJoCo 3.3.4.

Property Value
Base model lerobot/pi0
Training data hanapasta/airvla_v24 (FT-C list + 160 on-policy corrective episodes)
Initialised from FT-C
Training steps 20,000 (lr 5e-6 -> 5e-7); checkpoint 17,500 selected
Checkpoint selection lowest pinned-noise validation MSE over the saved ladder

Frozen-protocol result (n = 60 pick + 20 navigation, paired scenes)

1/60 grasped - 37/60 correct-target - 196.0 mm median - 82.0 mm target-true, the best pure-policy precision of the campaign - 12/20 navigation. The selection regression vs FT-C is significant (exact McNemar p = .013).

Usage

Evaluated with the frozen harness eval_v2.py from the code repository:

python eval_v2.py <this_checkpoint> 60 20 --torchseed 1000 --tag run --video

Code, reproduction guide and the full experimental ledger: https://github.com/robotics-hana/drone-version2

Evaluation uses MuJoCo 3.3.4; a different simulator version changes contact behaviour enough to invalidate comparison with the banked results.

Downloads last month
13
Safetensors
Model size
4B params
Tensor type
F32
·
BF16
·
Video Preview
loading