---
license: apache-2.0
task_categories:
- robotics
tags:
- LIBERO
- robosuite
- mujoco
- visual-domain-variation
---

# LIBERO Spatial Visual Variations

This derived dataset expands `libero_spatial` with six visual conditions per
source trajectory: the original LIBERO appearance plus five scene-level visual
palettes. It preserves the original task states, actions, rewards, dones,
robot state vectors, proprioceptive observations, task language, and BDDL task
semantics.

Visual conditions:

- `original`: Original - Default LIBERO spatial visual appearance.
- `nordic-sage`: Nordic Sage - White marble, pale plaster, light wood, sage cabinet, warm showroom light.
- `midnight-fjord`: Midnight Fjord - Dark floor, deep blue wall, gray table, navy cabinet, cool low-key light.
- `terracotta-pop`: Terracotta Pop - Brown ceramic, yellow linen, wood table, clay cabinet, warm terracotta light.
- `alpine-oak`: Alpine Oak - Wood planks, white wall, light wood table, white-oak cabinet, clear daylight.
- `graphite-gallery`: Graphite Gallery - Gray ceramic, dark plaster, gray table, charcoal cabinet, crisp spotlight.

The dataset contains ten LIBERO spatial tasks. Each task file contains the
first ten source demonstrations rendered under all six visual conditions, for
60 trajectories per task and 600 trajectories total.

Within every task HDF5, the trajectory blocks are ordered as:

```text
demo_0..demo_9    original
demo_10..demo_19  nordic-sage
demo_20..demo_29  midnight-fjord
demo_30..demo_39  terracotta-pop
demo_40..demo_49  alpine-oak
demo_50..demo_59  graphite-gallery
```

Prefer reading `visual_condition_key` and `source_demo_index` from each
episode's attributes instead of relying only on this index convention.

Each episode stores explicit metadata:

- `task_name`, `task_index`, `language_instruction`
- `source_demo_key`, `source_demo_index`, `source_task_file`
- `visual_condition_key`, `visual_condition_name`, `background_style`
- `camera_calibration` group with per-camera intrinsics, per-frame extrinsics,
  and per-frame world-to-pixel projection matrices

Common trajectory and observation keys follow the usual LIBERO HDF5 layout:

```text
data/demo_*/actions
data/demo_*/states
data/demo_*/robot_states
data/demo_*/rewards
data/demo_*/dones
data/demo_*/obs/agentview_rgb
data/demo_*/obs/eye_in_hand_rgb
data/demo_*/obs/joint_states
data/demo_*/obs/gripper_states
data/demo_*/obs/ee_pos
data/demo_*/obs/ee_ori
data/demo_*/obs/ee_states
```

Camera calibration is stored at:

```text
data/demo_*/camera_calibration/<camera_name>/intrinsic_matrix
data/demo_*/camera_calibration/<camera_name>/extrinsic_matrices
data/demo_*/camera_calibration/<camera_name>/world_to_pixel_matrices
```

Calibration is available for both `agentview` and `robot0_eye_in_hand`.
The array shapes are:

```text
intrinsic_matrix          (3, 3)
extrinsic_matrices        (T, 4, 4)
world_to_pixel_matrices   (T, 4, 4)
```

The uploaded dataset was re-downloaded from Hugging Face and validated for:

- 10 task files
- 600 total trajectories
- expected visual-condition blocks
- expected RGB observation shapes `(T, 128, 128, 3)`
- expected camera calibration groups and matrix shapes