airvla_v2_pi0_047500

AirVLA v2 baseline — pi0 checkpoint 047500

The T4-selected checkpoint of the AirVLA v2 study (SkyGrip drone + 2-DoF arm, MuJoCo): pi0 (4.03B) fine-tuned with LeRobot 0.6 on hanapasta/airvla_v2 (dataset revision 2e40a5c89746; 480-episode train split, seed 424242).

• Training: batch 4, seed 1000, bf16, grad checkpointing, vision encoder unfrozen; steps 30k -> extended to 60k; save_freq 2500.
• Selection: lowest teacher-forced validation action-MSE over the full 24-checkpoint curve (pre-registered rule): 047500, MSE 0.0000455 (bracketed minimum; 60k = 0.000055).
• Frozen closed-loop eval (v2 world, replay-validated harness, naive 50-step chunks, scene seed 97000 / torch seed 1000): pick n=60 median miss 170.8 mm, 1 picked, 1 placed (first complete pick-and-place of the project); nav 9/20 with 20/20 gate crossings.

Load

from lerobot.policies.pi0.modeling_pi0 import PI0Policy
policy = PI0Policy.from_pretrained("hanapasta/airvla_v2_pi0_047500")

Evaluate ONLY in the v2 world via eval_v2.py (repo: drone-version2, Sim'n'Real/Mujoco/) — the v1 harness renders a different world. Full study record: Reports/V2_STUDY.md in the project repo; per-run provenance in the eval logs' PROV lines.