qualiadev/pi05-smoke-openarm-pickv6

qualia_pi05 / smoke_openarm_pickv6

openpi pi05 SFT fine-tune, launched via q-research vlatrain.

• Dataset: qualiadev/openarm_pick_v6 (LeRobot v2.1)
• Steps: 500 | Batch: 32

Revisions step-<N> hold inference params only; the default revision is the final full checkpoint (params + train_state).

Run config (verbatim, for reproducibility)

# SMOKE run — validate the q-research train → openpi → HF pipeline end to end, not model quality.
# Dataset: qualiadev/openarm_pick_v6 — our copy of the public AiSaurabhPatil/openarm_pick_v6, a
# LeRobot **v2.1** OpenArm bimanual set (16-DOF: 7 joints + gripper per arm, 3 cams, 100 episodes).
# We re-hosted it under qualiadev and added the `v2.1` git tag: openpi's pinned lerobot
# (`get_safe_version`) rejects any repo with no codebase-version tag, and the upstream repo has none.

name: qualia_pi05                 # in-process registry name; leave as-is
exp_name: smoke_openarm_pickv6    # checkpoint dir: <checkpoint_base_dir>/<name>/<exp_name>/
repo_id: qualiadev/openarm_pick_v6   # LeRobot v2.1 dataset on the HF Hub (fetched on the VM), tagged v2.1
checkpoint_base_dir: /ephemeral/checkpoints  # vlatrain (remote) → the VM's big ephemeral disk
project_name: qualia-pi05         # wandb project (online when WANDB_API_KEY is set, else offline)
hf_repo_id: qualiadev/pi05-smoke-openarm-pickv6  # token = qualiadev (personal); it is in no org

# ── Data: OpenArm bimanual (generic layer) ──────────────────────────────────────────────────────────
data_layer: generic
cameras:                          # cam_1 -> base_0_rgb, cam_2/cam_3 -> left/right wrist
  cam_1: observation.images.head
  cam_2: observation.images.wrist_left
  cam_3: observation.images.wrist_right
relativize: true                  # mask auto-derived from action names (joints relative, gripper absolute)
bimanual: true                    # exactly 2 gripper dims (left_gripper, right_gripper) -> assert passes

# ── Model ──────────────────────────────────────────────────────────────────────────────────────────
action_horizon: 16                # chunk length; model action width stays 32 regardless of real DOF

# ── Hyperparameters (smoke-sized) ────────────────────────────────────────────────────────────────────
num_train_steps: 500              # smoke: just prove the loop + checkpointing run
save_interval: 250                # one intermediate save (step-250 params revision) + the final full ckpt
keep_period: null                 # keep only the latest local checkpoint
batch_size: 32
num_workers: 8
seed: 42
fsdp_devices: 1
ema_decay: null                   # EMA disabled
overwrite: false                  # set true to re-run into an existing exp_name
optimizer:                        # AdamW
  b1: 0.9
  b2: 0.95
  eps: 1.0e-8
  weight_decay: 1.0e-10
  clip_gradient_norm: 1.0
lr_schedule:                      # cosine decay with warmup, scaled to the 500-step smoke run
  warmup_steps: 50
  peak_lr: 2.5e-5
  decay_steps: 500                # match num_train_steps so the cosine lands at the floor at the end
  decay_lr: 2.5e-6