---
license: other
task_categories:
- robotics
- imitation-learning
pretty_name: UP-VLA Precision Recovery 15K
size_categories:
- 10K<n<100K
---

# UP-VLA Precision Recovery 15K

This research dataset contains 15,000 physically validated
precision-recovery samples in 150 WebDataset shards.

Each sample contains a wrist RGB image, metric depth, camera-frame XYZRGB point cloud,
and synchronized dual action trajectories:

1. relative end-effector SE(3) in the current tool frame;
2. robot joint positions on the same timestamps.

Perturbations are 5 mm to 5 cm and 5 to 30 degrees. Samples include deterministic
randomization receipts for background textures, clutter, lighting, RGB photometrics,
depth noise, and depth dropout.

## Coverage

- Unique object models: 111
- Source families: {"rlbench": 5220, "robotwin_generic": 9780}
- Task categories: {"edge_grasp": 1617, "insertion_alignment": 5200, "precision_grasp": 8183}
- Domain-randomized samples: 15000
- Background signatures: 5291
- Scene signatures: 5293
- Lighting signatures: 14659
- Sensor signatures: 14573
- All source experts successful: yes
- All recovery tasks successful: yes

## Format

Use manifest.jsonl to map sample keys to shards/shard-*.tar. Each key has
metadata.json, wrist_rgb.png, wrist_depth.npy, wrist_pointcloud.npy, and
action_chunk.npz members.

## License and source notice

RoboTwin-derived samples use the MIT-licensed RoboTwin software and assets subject to
their own provenance. RLBench-derived samples are for non-commercial internal or
academic research under the RLBench licence from Imperial College London. Publications
or presentations using those samples must identify RLBench as their source and comply
with the upstream licence. This combined dataset is therefore distributed for
non-commercial research use only.

The dataset contains rendered observations and trajectories, not copies of simulator
source code.