sai-prasanna/robotwin-2.0-480p

RoboTwin 2.0 Clean-50 at 480p

This dataset contains 2,500 rerendered RoboTwin ALOHA-AgileX demonstrations: 50 tasks with 50 episodes per task. The head and left and right wrist cameras use native 640×480 rendering. The auxiliary front camera retains its native 320×240 resolution. Images are JPEG-encoded; they are not upscaled.

The source demonstrations come from TianxingChen/RoboTwin2.0. Rerendering uses RoboTwin revision 6dde57155eafa3e4ebf6ad1f93a7cf7d5d41a755 and the Large_D435 camera profile for the head and wrist cameras. The collection, conversion, and publication scripts are included in code/ for reproducibility.

Download a task

Each task follows the official Clean-50 ZIP naming and directory layout. The core files are:

dataset/TASK/aloha-agilex_clean_50.zip
└── aloha-agilex_clean_50/
    ├── data/episode0.hdf5 … episode49.hdf5
    ├── instructions/episode0.json … episode49.json
    ├── video/episode0.mp4 … episode49.mp4
    ├── _traj_data/episode0.pkl … episode49.pkl
    ├── seed.txt
    └── scene_info.json

The archives also include per-task provenance and replay-check logs.

To download and extract one task, run:

from pathlib import Path
from zipfile import ZipFile
from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="sai-prasanna/robotwin-2.0-480p",
    repo_type="dataset",
    filename="dataset/adjust_bottle/aloha-agilex_clean_50.zip",
)
with ZipFile(archive) as source:
    source.extractall(Path("data/adjust_bottle"))

HDF5 schema

The export maps the rerendered XPolicyLab observations to the official RoboTwin Clean-50 HDF5 keys. JPEG bytes, observed poses, and observed joint states retain their native values. Episode filenames use episodeN without zero padding.

Key	Contents
observation/head_camera/rgb	JPEG frames at 640×480
observation/left_camera/rgb	Left wrist JPEG frames at 640×480
observation/right_camera/rgb	Right wrist JPEG frames at 640×480
observation/front_camera/rgb	Auxiliary front JPEG frames at 320×240
observation/CAMERA/intrinsic_cv	Per-frame 3×3 camera intrinsics
observation/CAMERA/cam2world_gl	Per-frame 4×4 OpenGL camera-to-world matrices
observation/CAMERA/extrinsic_cv	Per-frame 3×4 OpenCV world-to-camera matrices
endpose/left_endpose, endpose/right_endpose	Observed XYZ and WXYZ quaternion, 7 values per arm
endpose/left_gripper, endpose/right_gripper	Observed scalar gripper values
joint_action/left_arm, joint_action/right_arm	Observed 6-joint arm states
joint_action/left_gripper, joint_action/right_gripper	Observed scalar gripper values
joint_action/vector	Left arm, left gripper, right arm, right gripper; 14 values
pointcloud	Empty per-frame point clouds, shape (T, 0)

Additional action, instructions, additional_info, and data_format_version entries preserve native commanded actions and collection metadata. action contains native commands; joint_action contains observed states, as in the base RoboTwin schema. Camera extrinsics are converted from OpenGL to OpenCV coordinates using diag(1, -1, -1, 1) @ inverse(cam2world_gl).

Validation

The release contains 50 task archives and 2,500 episodes, with 536,464 frames per camera. Validation fully decoded all 2,145,856 JPEGs and checked each camera's resolution. It also checked episode counts, finite pose and joint values, frame/state alignment, camera calibration shapes, and every ZIP's CRC values.

manifest.json records per-task episode counts, decoded JPEG counts, audited repairs, archive sizes, and SHA-256 checksums. SHA256SUMS provides the same archive checksums for command-line verification. Every uploaded archive's size and SHA-256 were verified against the release manifest.

Audited replay repairs

Every episode retains its corresponding original source seed and object identities. No episode uses another demonstration as a donor.

The audit identified 22 replays that failed the simulator's task success check and 202 episodes whose object variants differed from the official source. Three episodes belonged to both groups, giving 221 affected episodes across 22 tasks. The object mismatch came from filesystem-dependent ordering of model files. The rerenderer restores the model IDs recorded in the official scene_info.json while preserving the seeded random draws.

The 221 corrected episodes pass the unmodified simulator success predicate and remain successful at 0.25, 0.5, 1, and 2 seconds afterward. No terminal success check is bypassed. Repairs retain recorded arm paths where possible; contact-sensitive failures use settling time, preserved source control flow, or explicitly logged contact-path replanning. These are derived replays, and some motions and episode lengths differ from the official videos.

For put_bottles_dustbin episode 28, the first two bottles retain their source motions. The final bottle uses a reachable neck grasp and a table handoff within the left arm's workspace before disposal. This repairs the physical pickup instead of accepting a failed handoff.

Episodes 30 and 31 of put_bottles_dustbin have separate audited corrections:

• Episode 30 uses its recovered original replay. Its right arm stays stationary, matching the official video and joint states. A collection runner had incorrectly quarantined this completed episode when episode 31 failed.
• Episode 31 preserves the source's left–right–left arm choices. Small object motion differences crossed an arm-selection boundary during replay. Preserving the recorded choices replays all original paths and passes the simulator's terminal success check.

Episodes 30 and 31 retain all original planned paths. Their observed arm joints match the source frames exactly; gripper differences are limited to floating-point rounding, and the largest end-effector position component difference is 0.37 mm. Their earlier donor substitutions are removed.

Across the 221 corrected episodes, 212 retain every original planned arm path. The comparison verifies 1,470 unchanged segments and records 29 replanned segments.

repair_plan.json records the selected repair configurations and candidate runs. To reproduce corrected episodes, use those configurations and the provided repair scripts: upstream replay alone does not include added settling or handoff changes. comparison.json records source and exported trajectory hashes and identifies every changed arm-path segment. Per-task provenance includes the success and stability results. Superseded failure events remain in a separate history field; they do not describe accepted failures in this release.

The release preserves rerendered frames and states without padding or replacing them with the source release's lower-resolution observations. Physics replay is not pixel-identical to the source videos.