fanqi-robo/pi05_busybox_push_green_button_expert_only

pi05 · BusyBox "push the green button" · variant expert_only

Ablation arm expert_only (group trainable_params) of the π0.5 fine-tuning study in experimental/lerobot_policy_pi05/ (alpha-robotics, branch policy_test_fanqi).

• Description: b0 with the whole PaliGemma VLM frozen; only the 300M action expert and the state/action projections train (LeRobot train_expert_only).
• Hypothesis: Dream Machines found LoRA (even rank 256) 23+ pp behind a full fine-tune and argue a new embodiment is not a low-rank edit. Expert-only is the other cheap regime (fits a 24-32 GB card, ~2x faster) and was not measured there. If it is close to b0 on this task we can move future sweeps to cheap Vast 5090s; if it is far behind, that settles the "full fine-tune only" question for us as well.
• Base weights: lerobot/pi05_base @ a538eb27 (LeRobot 0.5.1, --policy.pretrained_path; pinned because later revisions of the base use a processor step 0.5.1 cannot load)
• Dataset: armnet/busybox_push_green_button — all 39 episodes (no hold-out; robot eval only)
• Published checkpoint: step 10000 (root of this repo, tag step-10000)
• Earlier checkpoints: fanqi-robo/pi05_busybox_push_green_button_expert_only_ckpts (checkpoints/<step>/pretrained_model, newest 2); full history on the Modal Volume pi05-busybox-ckpts
• W&B: https://wandb.ai/fanqi-robo-saferobotics/pi05_busybox_push_green_button

Evaluate

The only test is the real robot: submit fanqi-robo/pi05_busybox_push_green_button_expert_only on https://huggingface.co/spaces/armnet/armnet-eval (embodiment lerobot/so-101, task push_green_button, 20 rollouts, variation seed 42). There is no held-out split. Loading needs lerobot[pi] and a Hub token that can read the gated google/paligemma-3b-pt-224 tokenizer.

Training config (lerobot-train --config_path=...)

policy:
  type: pi05
  pretrained_path: lerobot/pi05_base
  device: cuda
  dtype: bfloat16
  gradient_checkpointing: true
  compile_model: false
  chunk_size: 50
  n_action_steps: 50
  num_inference_steps: 10
  tokenizer_max_length: 200
  freeze_vision_encoder: false
  train_expert_only: true
  use_relative_actions: false
  normalization_mapping:
    VISUAL: IDENTITY
    STATE: QUANTILES
    ACTION: QUANTILES
  optimizer_lr: 2.5e-05
  optimizer_weight_decay: 0.01
  optimizer_grad_clip_norm: 1.0
  scheduler_warmup_steps: 1000
  scheduler_decay_steps: 30000
  scheduler_decay_lr: 2.5e-06
  push_to_hub: false
  repo_id: fanqi-robo/pi05_busybox_push_green_button_expert_only
dataset:
  repo_id: armnet/busybox_push_green_button
  episodes:
  - 0
  - 1
  - 2
  - 3
  - 4
  - 5
  - 6
  - 7
  - 8
  - 9
  - 10
  - 11
  - 12
  - 13
  - 14
  - 15
  - 16
  - 17
  - 18
  - 19
  - 20
  - 21
  - 22
  - 23
  - 24
  - 25
  - 26
  - 27
  - 28
  - 29
  - 30
  - 31
  - 32
  - 33
  - 34
  - 35
  - 36
  - 37
  - 38
  image_transforms:
    enable: false
batch_size: 32
num_workers: 8
steps: 10000
save_freq: 2000
eval_freq: -1
log_freq: 20
seed: 1000
job_name: pi05_bb_green_expert_only
output_dir: /root/outputs/train/pi05_bb_green_expert_only
wandb:
  enable: true
  entity: fanqi-robo-saferobotics
  project: pi05_busybox_push_green_button
  mode: online
  disable_artifact: true