RawVLA-Bench

RawVLA-Bench is a paired training-trajectory dataset for studying RAW-domain visual frontends for vision-language-action (VLA) policies. It is derived from successful expert trajectories in LIBERO and RoboTwin 2.0.

Each trajectory has a RAW input version captured or rendered under a sampled lighting condition and a paired default-light RGB target. The pair shares the same trajectory identity, actions, robot states, frame count, and relative path. Failed trajectories and replay-debug artifacts are not included.

Five-light RAW preview

LIBERO

The grid shows middle frames from five successful LIBERO trajectories for the same task. Rows are agent-view and wrist-view RAW inputs; columns are the five lighting domains and report the sampled exposure value. Pixels use the stored sensor-domain uint8 values directly, without per-image normalization or brightness enhancement, so the low-light inputs intentionally appear dark.

RoboTwin 2.0

This grid uses five successful grab_roller trajectories. Rows are the agent, left-wrist, and right-wrist RAW inputs; columns are the five lighting domains. The stored linear float32 [0,1] values are clipped to that fixed range and mapped to uint8 only for display, with no per-image normalization or brightness enhancement.

Dataset contents

Source	Successful paired trajectories	Views	RAW representation
LIBERO	1,771	agent view, wrist	uint8, 3-channel sensor-domain representation
RoboTwin 2.0	632	head, left, right	float32 [0,1], 3-channel linear sensor-domain representation
Total	2,403

There are 4,806 NPZ files: one RAW file and one RGB target file for each of the 2,403 successful trajectories.

data/
├── libero/
│   ├── raw/<suite>/episode_XXXXXX/variant_00.npz
│   └── rgb/<suite>/episode_XXXXXX/variant_00.npz
└── robotwin2/
    ├── raw/<task>/episode_XXXXXX.npz
    └── rgb/<task>/episode_XXXXXX.npz

Pair files by their path relative to raw/ or rgb/.

LIBERO subset

The LIBERO subset covers libero_spatial, libero_object, libero_goal, and libero_10. Demonstrations were replayed from the per-demonstration model XML and HDF5 initial state, with the standard 10 dummy steps and OpenVLA no-op filtering. The RAW and default-light RGB versions use the same saved actions, states, frame selection, and camera convention.

All 1,771 published episodes reached success in the original OpenVLA/OFT replay-generation environment. A separate audit in a newer StarVLA simulator environment reproduced 1,474 successes; the other 297 were rechecked in the generation environment and succeeded, indicating simulator-version drift rather than failed generated trajectories.

LIBERO RAW files contain:

• agentview_raw_uint8, wrist_raw_uint8
• actions, states, robot_states, rewards, dones
• metadata_json

The corresponding RGB files replace the image fields with agentview_rgb_uint8 and wrist_rgb_uint8.

RoboTwin 2.0 subset

The RoboTwin subset begins from the strict success-only allowlist produced by clean closed-loop expert replay. Inclusion required a completed replay, success=true, an existing replay HDF5 file, available PSNR/SSIM metrics, and matching original/replay frame counts for all three cameras.

The allowlist contains 633 successful episodes across 13 tasks. One allowlisted episode, put_bottles_dustbin/episode_000031, failed during paired lighting export and is omitted. The published subset therefore contains 632 complete RAW/RGB pairs and no task-failure episodes.

For RAW generation, the clean successful replay's physics state was copied at each frame and rendered under the sampled lighting condition without advancing the perturbed environment with actions. This preserves alignment with the successful clean trajectory.

RoboTwin RAW files contain:

• head_camera_raw_float32, left_camera_raw_float32, right_camera_raw_float32
• actions, robot_states, left_endpose, right_endpose, rewards, dones
• metadata_json

The corresponding RGB files replace the image fields with head_camera_rgb_uint8, left_camera_rgb_uint8, and right_camera_rgb_uint8.

Loading an episode

import numpy as np

raw_path = "data/libero/raw/libero_spatial_no_noops_1.0.0_lerobot/episode_000000/variant_00.npz"
rgb_path = raw_path.replace("/raw/", "/rgb/")

with np.load(raw_path, allow_pickle=False) as raw, np.load(rgb_path, allow_pickle=False) as rgb:
    actions = raw["actions"]
    raw_frames = raw["agentview_raw_uint8"]
    target_frames = rgb["agentview_rgb_uint8"]

NPZ files are episode-level and may be large. Load one episode at a time or use memory-aware preprocessing when constructing frame or sequence datasets.

Intended use

RawVLA-Bench is intended for research on RAW-to-RGB or RAW-to-feature frontends, illumination robustness, paired representation learning, and VLA policy training. It contains simulation data and should not be interpreted as evidence of real-world robot safety or deployment readiness.

This release contains training trajectories only. It does not define a held-out test split or a leaderboard metric.

Licenses and attribution

This is a mixed-license derived dataset:

• The LIBERO-derived subset retains the upstream Creative Commons Attribution 4.0 International (CC BY 4.0) dataset license and attribution requirements.
• The RoboTwin 2.0-derived subset follows the upstream MIT license.

Users are responsible for complying with the applicable upstream license for each subset. Please also cite the corresponding source benchmark.

Citations

@article{liu2023libero,
  title={LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning},
  author={Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and Liu, Qiang and Zhu, Yuke and Stone, Peter},
  journal={arXiv preprint arXiv:2306.03310},
  year={2023}
}

@article{chen2025robotwin,
  title={RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation},
  author={Chen, Tianxing and Chen, Zanxin and Chen, Baijun and Cai, Zijian and Liu, Yibin and others},
  journal={arXiv preprint arXiv:2506.18088},
  year={2025}
}

Upstream resources

• LIBERO: https://libero-project.github.io/
• LIBERO demonstrations: https://huggingface.co/datasets/yifengzhu-hf/LIBERO-datasets
• RoboTwin 2.0: https://robotwin-platform.github.io/
• RoboTwin 2.0 dataset: https://huggingface.co/datasets/TianxingChen/RoboTwin2.0