robocasa365.yaml

type: robocasa365

# Umbrella directory holding the independent native-delta RoboCasa365 LeRobot
# v3 conversions (atomic + composite) one level down; they share one schema and
# are read as one 300-task collection.
dataset_dir: /path/to/robocasa365_data

# ── Action space ─────────────────────────────────────────────────────────────
# The converted Parquet columns are row-aligned:
#   observation.state[19] = eef xyz3 + rot6d6 + gripper1 + base xyz3 + rot6d6
#   action[15] = native normalized delta xyz3
#                + rot6d(Exp(native normalized delta rotvec3))
#                + flipped gripper1
#                + native base vx/vy/vyaw + torso + control_mode
# The rotation command is not multiplied by 0.5 and is not composed with state.
# State and action have different raw widths and independent normalization
# blocks — hence the separate unify_state_map. There is no 20-D intermediate.
action_mode: eef
unify_action: true
unify_action_map: ["0-9", "68-72"]
unify_state_map: ["0-9", "68-76"]

# ── Frame sampling ───────────────────────────────────────────────────────────
# num_frames = state/action window length; action horizon = num_frames - 1.
# video_stride subsamples video only; state/action stay at the raw rate.
# window_stride = sliding window step (1 = every frame starts a window).
# One raw state per sample and a 32-step same-row action horizon.
num_frames: 33
video_stride: 4
window_stride: 1
split: train

# ── Rendering (3-slot L-shape canvas) ────────────────────────────────────────
# camera_layout = [head, left/bottom-left, right/bottom-right]; null slots are
# rendered as black tiles. At 384x320: top 256Hx320W, bottom halves 128Hx160W,
# each tile BILINEAR-stretched to its slot. target_camera drives single-view
# mode and is ignored when multiview=true.
height: 384
width: 320
multiview: true
camera_layout:
  - observation.images.robot0_agentview_left
  - observation.images.robot0_eye_in_hand
  - observation.images.robot0_agentview_right
target_camera: observation.images.robot0_agentview_left

# ── Normalization ────────────────────────────────────────────────────────────
# Pooled stats live at <dataset_dir>/meta/robocasa365_normalization_stats.npy
# and are auto-built on first use (rank 0 computes, other ranks wait).
normalize_mode: min-max
normalization_stats_path: null  # override only; default auto-resolves under <dataset_dir>/meta/

# ── Video augmentation (optional) ────────────────────────────────────────────
# Load-time color jitter applied by the reader to each clip with the SAME
# random factors across all frames (temporal consistency), TRAIN split only.
# enabled: false keeps video byte-identical while preserving the knobs.
color_jitter:
  enabled: true
  brightness: 0.2
  contrast: 0.2
  saturation: 0.2
  hue: 0.0