Robotics
LeRobot
Safetensors
pi05
vla
yam

pi05_red_block_only

Full-parameter PI0.5 fine-tune on YzyLmc/red_block_only (47 episodes, 20,822 frames, 30 fps, task "Pick up the red block").

Corrected global-percentile normalization stats; action chunk 60 steps (2.0 s)..

This repo always holds the LATEST checkpoint of a run in progress.

Training step 002000 / 20000
Base lerobot/pi05_base @ a538eb273274
Batch / precision 32, bfloat16, gradient checkpointing on
LR 2.5e-5 peak, cosine to 2.5e-6, 1k warmup
Action 7-D, chunk of 50 @ 30 fps

Not evaluated on hardware. No held-out split: all 47 episodes are trained on, so any training-set metric reflects fit, not generalisation.

Cameras must be renamed onto the openpi-canonical keys at inference: scene_camera -> base_0_rgb, left_wrist_camera -> left_wrist_0_rgb.

config.json is corrected to declare a 7-D action; lerobot/pi05_base ships openpi's [32] default, which LeRobot does not override from the dataset. Weights are byte-identical to the trainer output.

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

Model tree for YzyLmc/pi05_red_block_only_v2

Finetuned
(529)
this model

Dataset used to train YzyLmc/pi05_red_block_only_v2