libero.yaml

type: libero

# Canonical LIBERO dataset. The reader accepts only the native-action
# conversion metadata (meta/conversion.json).
dataset_dir: /path/to/libero_data

# ── Action space ─────────────────────────────────────────────────────────────
# Normalize action and achieved state from their separate native-action stats
# blocks, then scatter both 10-D streams into the canonical left-arm slots.
# Rot6d slots (3:9) are always identity-mapped by the canonical reader; this
# is enforced even if a stale/custom stats file contains non-identity values.
# The benchmark bridge decodes the model's native 10-D action to LIBERO's
# runtime 7-D OSC command without absolute-goal composition or controller
# scale.
action_mode: eef
unify_action: true
unify_action_map: ["0-9"]
unify_state_map: null  # null = state reuses unify_action_map

# ── 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).
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.image, observation.images.image2, null]
target_camera: observation.images.image

# ── Normalization ────────────────────────────────────────────────────────────
# Stats live at <dataset_dir>/meta/libero_normalization_stats.npy and are
# auto-built on first use (rank 0 scans, other ranks wait).
normalize_mode: min-max

# ── 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: false
  brightness: 0.2
  contrast: 0.2
  saturation: 0.2
  hue: 0.0