pi0 fine-tuned for button-press (RM65-6F sim) โ€” step 150000

Fine-tuned from lerobot/pi0_base with train_expert_only=true on the local/button_press_0901 dataset (61 episodes, 30981 frames, 2 cameras side_view/wrist_ego, 6-dim state/action).

This is the final fine-tune: continued from the earlier checkpoints to step 150000 (~19.4 epochs, batch_size=4, bf16), with a cosine LR schedule decaying 2.5e-5 -> 2.5e-6 across the full run. Loss converged to a plateau (~0.03-0.04) by the end. It supersedes the previously uploaded step-15000 policy.

Controls the RM65-6F + DexHand-021 sim to press the red button. Task: "press the red button on the box".

Downloads last month
36
Safetensors
Model size
4B params
Tensor type
F32
ยท
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for yixiaosz/pi0_button_press

Base model

lerobot/pi0_base
Finetuned
(64)
this model