pi05_rlt_busybox_push_green_button

RL-token encoder/decoder (RLT Stage 1) trained on the frozen π0.5 green-button checkpoint. This is the checkpoint hw_control.pi0_rlt loads to extract tokens for the demo cache and online RL.

Policy	π0.5 + RLT (Pi0RLConfig, pi05=true)
Frozen VLA	pravsels/pi05_busybox_push_green_button
Dataset	villekuosmanen/busybox_push_green_button
Task	busybox_push_green_button
Prompt	push the green button
Action dim	6 (single-arm; 5 joints delta, gripper absolute)
Cameras	top → base_0_rgb, wrist → left_wrist_0_rgb, front → base_1_rgb
Training	GCloud 1× A100 80GB, VLA frozen (rl_vla_loss_weight=0.0), 20k steps, global batch 16, action horizon 30
W&B project	busybox_push_green_button_rlt
W&B run	wnk0bxds

Checkpoints

main is the finalized 20k checkpoint (params/ + assets/ at the repo root). train_state/ is not published.

Field	Value
Latest published	main (params/, step 19999)
Config	pi05_rlt_busybox_push_green_button
Code	4354d7d on task/rlt_busybox_green_button
Runtime	4h 15m (2026-08-27 20:04 UTC → 2026-08-28 00:19 UTC)
Train loss	9935.29 (step 0) → 189.18 (step 19900)

Usage

from huggingface_hub import snapshot_download
from openpi.policies import policy_config
from openpi.training.config import get_config

ckpt = snapshot_download("pravsels/pi05_rlt_busybox_push_green_button")
policy = policy_config.create_trained_policy(
    get_config("pi05_rlt_busybox_push_green_button"),
    ckpt,
)

uv run scripts/serve_policy.py \
  --policy.config=pi05_rlt_busybox_push_green_button \
  --policy.dir=/path/to/pravsels/pi05_rlt_busybox_push_green_button