pi0.5 LoRA fine-tune — pick_color_cube (Isaac Lab)
LoRA fine-tune of pi0.5 (LIBERO base) on Isaac Lab joystick demonstrations of color-cube pickup with a Franka arm. Used as the VLA backbone in a VLA-BCI shared-autonomy stack.
| Field | Value |
|---|---|
| openpi train config | pi05_pick_color_cube |
| base weights | gs://openpi-assets/checkpoints/pi05_libero/params |
| training dataset | hongyi/pick_color_cube (LeRobot v2.0, joystick demos) |
| step | 4999 (final of 5000-step run) |
| wandb run id | kh36rwlo |
| trained | 2026-04-09 |
| LoRA target | gemma_2b_lora + gemma_300m_lora action expert |
| action dims | 7D delta EE — [dx, dy, dz, 0, 0, 0, gripper] |
| action horizon | 10 |
VLA_POS_SCALE=1.0 is required at inference (vs the 0.02 used for the
zero-shot LIBERO checkpoint) because actions are already deltas in the
training data.
Serving
# 1) Download into the openpi checkpoint layout
huggingface-cli download hongyyyy/pi05_pick_color_cube_v2 \
--local-dir pi_policy/openpi_vela/checkpoints/pi05_pick_color_cube/pick_cube_pi05_lora_v2/4999
# 2) Serve
cd pi_policy/openpi_vela
uv run python scripts/serve_policy.py policy:checkpoint \
--policy.config pi05_pick_color_cube \
--policy.dir checkpoints/pi05_pick_color_cube/pick_cube_pi05_lora_v2/4999
Contents
params/— Orbax-sharded JAX params (Pi0 + LoRA adapters)assets/hongyi/pick_color_cube/norm_stats.json— action/state normalisation_CHECKPOINT_METADATA— Orbax metadata
train_state/ (the optimizer state, ~2.9 GB) is not included — it is only
needed to resume training, not to serve. Resume training from the original
machine.