training_code/data_config.yaml

# Data config: fruit_pick, 100 real-robot demos, Franka panda, 2-cam.
# Downloaded 2026-08-09 from
# https://huggingface.co/datasets/aabyaneh/fruit_pick -- LeRobot codebase_version
# v3.0 (file-chunked, NOT the v2.1 per-episode-file layout our own
# lift2lerobot converter produces), confirmed the vendored lerobot loader here
# (fastwam/datasets/lerobot/lerobot/lerobot_dataset.py) supports v3.0 natively,
# so no reconversion needed.
#
# SCHEMA VERIFIED AGAINST meta/info.json (2026-08-09), differs from every
# other dataset in this repo in one way:
#   - camera keys are "agentview"/"wrist" (not "image"/"wrist_image")
#
# observation.state came off HF as a native 15-dim vector (eef_xyz[3] +
# eef_quat_xyzw[4] + joint_0..6[7] + gripper_width[1]) -- NOT the 8-dim
# eef_pose+gripper convention every other dataset here uses. Per user
# request (this run must match place_cube_in_bowl/lift_new exactly),
# fastwam_train/scripts/fix_fruit_pick_state_schema.py rewrote the LOCAL
# parquet's observation.state down to the SAME 8-dim formula
# lift2lerobot/convert_lift_hdf5_to_lerobot_v21.py uses: eef_pos(3) +
# quat2axisangle(eef_quat)(3) + [gripper_width/2, -gripper_width/2] --
# bit-identical convention, not a from-scratch reimplementation. Original
# 15-dim parquet backed up to data/chunk-000/file-000.parquet.orig_15dim_state.
#
# action is 7-dim (delta_xyz, delta_rot_xyz, grasp), SAME as every other
# task's schema/mask already; delta_rot_x/y/z are exactly 0.0 across all
# 20,149 frames (verified), matching the no-rotation convention used
# everywhere else in this pipeline. No changes needed there.
#
# Task prompt: "Lift the lid, put it aside, and pick the black plum." -- the
# dataset's own meta/tasks.parquet as downloaded said "pick up the fruit and
# place it in the basket" instead; user chose to keep their own prompt, so the
# LOCAL copy of data/fruit_pick/meta/tasks.parquet was overwritten to match
# (original backed up to tasks.parquet.orig_hf_label) -- the HF repo itself
# still has the old label. Must match the persistent T5 text cache exactly
# (see fastwam_train/scripts/precompute_text_embeds.sh).
#
# val_set_proportion 0.0 -- NO validation (user decision 2026-08-09: use the
# entire 100-episode dataset for training, no held-out split). No
# `episode_indices` either -- the full 100 episodes are used as-is; there's
# no larger pool to subset from (unlike George's pcbnew100, which slices 100
# out of place_cube_new's 500), so no capping is needed.
#
# LOCAL-DISK paths: dataset_dirs and text_embedding_cache_dir point at /dev/shm
# copies staged by fastwam_train/scripts/stage_fruit_pick_shm.sh -- stage
# BEFORE launching; re-stage after any node reboot. Persistent copy:
# data/fruit_pick/ (LeRobot v3.0, 100 episodes, 20,149 frames).
train:
  _target_: fastwam.datasets.lerobot.robot_video_dataset.RobotVideoDataset
  dataset_dirs:
    - /dev/shm/fruit_pick_lerobot_v21
  shape_meta:
    images:
      - key: agentview
        raw_shape: [3, 256, 256]
        shape: [3, 224, 224]
      - key: wrist
        raw_shape: [3, 256, 256]
        shape: [3, 224, 224]
    action:
      - key: default
        raw_shape: 7 # before transform
        shape: 7 # after transform
    state:
      - key: default
        raw_shape: 8 # before transform -- eef_pos(3)+axisangle(3)+gripper(2), see header
        shape: 8 # after transform
  num_frames: 33
  global_sample_stride: 1
  action_video_freq_ratio: 4 # 32 action, 9 video frames
  video_size: [224, 448] # final resize video
  camera_key: null
  val_set_proportion: 0.0
  is_training_set: true
  skip_padding_as_possible: false
  concat_multi_camera: "horizontal"
  processor:
    _target_: fastwam.datasets.lerobot.processors.fastwam_processor.FastWAMProcessor
    shape_meta: ${data.train.shape_meta}
    num_obs_steps: ${data.train.num_frames}
    num_output_cameras: 2
    action_output_dim: 7 # eef delta xyz (3) + delta rot xyz (3, always 0) + grasp (1)
    proprio_output_dim: 8 # eef_pos(3) + axisangle(3) + gripper(2) -- matches place_cube/lift_new
    delta_action_dim_mask:
      default: [true, true, true, true, true, true, false] # eef poses are delta, gripper is not

    # action & state normalization
    action_state_transforms: null
    use_stepwise_action_norm: False
    norm_default_mode: min/max
    norm_exception_mode: null

    action_state_merger:
      _target_: fastwam.datasets.lerobot.transforms.action_state_merger.ConcatLeftAlign

    train_transforms:
      - _target_: fastwam.datasets.lerobot.transforms.image.ToTensor
      - _target_: torchvision.transforms.Resize
        size: [224, 224]
    val_transforms:
      - _target_: fastwam.datasets.lerobot.transforms.image.ToTensor
      - _target_: torchvision.transforms.Resize
        size: [224, 224]
  text_embedding_cache_dir: /dev/shm/fruit_pick_text_embeds_cache
  context_len: 128