pi05_rlt_busybox_multitask_singlearm

RL-token encoder/decoder (RLT Stage 1) on the frozen prompt-fix π0.5 BusyBox multitask checkpoint. This is what hw_control.pi0_rlt loads to extract tokens.

Same relative 6D recipe as the VLA (pravsels/pi05_busybox_multitask): 5 joints delta, gripper absolute, per-timestep 1%/99%. Prompts are remapped per frame (prompt_from_task, RCW git main #597aa9ad).

Policy	π0.5 + RLT (Pi0RLConfig, pi05=true)
Frozen VLA	pravsels/pi05_busybox_multitask (W&B 4ym0qegc)
Dataset	villekuosmanen/busybox_multitask (66 episodes, 27 tasks)
Prompt	prompt_from_task (27 remapped instructions; no single default)
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
Norm	per-timestep 1%/99% (Hub assets copied, not recomputed)
W&B project	busybox_multitask_rlt_singlearm
W&B run	xmkdxvrl

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_multitask_singlearm
Code	29fb3dc on task/rlt_busybox_multitask
Runtime	4h 10m (2026-09-03 10:46 UTC → 14:56 UTC)
Train loss	12117.80 (step 0) → 246.70 (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_multitask_singlearm")
policy = policy_config.create_trained_policy(
    get_config("pi05_rlt_busybox_multitask_singlearm"),
    ckpt,
)

uv run scripts/serve_policy.py \
  --policy.config=pi05_rlt_busybox_multitask_singlearm \
  --policy.dir=/path/to/pravsels/pi05_rlt_busybox_multitask_singlearm