YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
pi0.5 on tmhoanggg/so101-stack-cup-v4
Fine-tuned from lerobot/pi05_base on tmhoanggg/so101-stack-cup-v4 with LeRobot (commit f5c9bc02), same recipe for every row:
vision encoder frozen, Gemma + action expert trained, 30k steps x batch 64, lr 2.5e-5 -> 2.5e-6, warmup 1k,
absolute joints, chunk 50, quantile normalisation. Cameras: front + wrist. Episode 103 (6 frames) left out.
One folder per checkpoint (every 5k steps), named <policy>_<k>k.
| policy | vision | rollout env |
|---|---|---|
official_frozen_* |
pi0.5's own SigLIP (frozen) | none |
virgil_vit_* |
VIRGIL (frozen) swapped in for SigLIP, new projector | PI05_ENCODER=virgil_vit |
dinov2_* |
DINOv2-base (frozen) swapped in for SigLIP, new projector | PI05_ENCODER=dinov2 |
option3_virgil_pred_* |
MAIN METHOD encoder (L_pred + delta L_sem, LIBERO 55.08) in the option-3 side branch | PI05_AUX=virgil_pred |
option3_<enc>_* (stack-cup only) |
SigLIP + side branch with the encoder; pi0.5 AND branch trained, from pi05_base | PI05_AUX=<enc> |
dynaflip_<enc>_* (stack-cup only) |
DynaFLIP App. D.4: pi05_base frozen + side branch; 10k x 32, lr 1.5e-5 | PI05_AUX=<enc> |
virgil_vit_swapv2_* (stack-cup only) |
VIRGIL swap, two stages: 2k steps projector only (lr 1e-4), then the recipe; image augmentation | PI05_ENCODER=virgil_vit |
official_frozen_relative_* (drawer only) |
official, RELATIVE joint actions (gripper absolute) | none |
<enc> = virgil_vit or dinov2.
Code: code/ (lerobot_rollout_pi05.py, encoder_swap.py, aux_inject.py). VIRGIL and DINOv2 need
github.com/tmhoanggg/r3m (main) on PYTHONPATH. Use the same n_action_steps and fps for every policy.
PI05_ENCODER=virgil_vit python code/lerobot_rollout_pi05.py <lerobot-rollout args> --policy.path=<folder>
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support