pi05_busybox_multitask_abs

Full-component fine-tune of π0.5 on villekuosmanen/busybox_multitask. Absolute 6D actions with per-timestep min/max bounds mapped to [-1, 1] and per-episode task prompts.

The relative-action sibling is pravsels/pi05_busybox_multitask.

Policy	π0.5 (pi05=true, full-component)
Init checkpoint	pi05_base (gs://openpi-assets/checkpoints/pi05_base)
Dataset	villekuosmanen/busybox_multitask (LeRobot v3, 66 episodes, 12141 frames, 20 fps, 27 tasks)
Prompt	prompt_from_task (no single default)
Action dim	6 (single-arm; absolute joints + gripper)
Cameras	top → base_0_rgb, wrist → left_wrist_0_rgb, front → base_1_rgb
Training	Vast 4× H100 80GB SXM data-parallel, full-component, 30k steps, global batch 32, action horizon 30
Norm	per-timestep min/max stored in q01/q99 (use_min_max_norm_stats=True)
W&B project	busybox_multitask_pi05_abs
W&B run	1xb73qsx

Checkpoints

main is the latest finalized checkpoint (params/ + assets/ at the repo root). The publisher overwrites the root at each 5k save. Training target is step 30,000. train_state/ is not published. Assets are from this config's own stats dir; they are not the relative-run stats.

Field	Value
Latest published	main (params/, step 29999)
Config	pi05_busybox_multitask_abs
Code	e846282 on task/busybox_multitask
Runtime	~4h 25m (2026-09-01 23:18 UTC → 2026-09-02 03:43 UTC)
Train loss	0.2664 (step 0) → 0.0017 (step 29900)

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_busybox_multitask_abs")
policy = policy_config.create_trained_policy(
    get_config("pi05_busybox_multitask_abs"),
    ckpt,
)

uv run scripts/serve_policy.py \
  --policy.config=pi05_busybox_multitask_abs \
  --policy.dir=/path/to/pravsels/pi05_busybox_multitask_abs