han-xudong/troncamp-mani-dataset

TronCamp Mani Expert Demonstrations

Raw RoboTwin expert demonstrations for the four 2026 TronCamp Mani tasks on the Tron2 dual-arm robot. The companion ACT checkpoints were trained from these demonstrations using the processing code in the GitHub repository.

Dataset Summary

Track	Task	Episodes	Frames, min/mean/max	Approx. size
T1	adjust_bottle	200	185 / 212.4 / 249	2.7 GB
T2	grab_roller	993	111 / 117.3 / 131	6.2 GB
T3	stack_bowls_two	1000	441 / 531.9 / 623	28 GB
T4	stack_bowls_three	1000	656 / 707.0 / 764	41 GB

The online ACT observation contract uses three RGB cameras and 16-D qpos. Raw files also contain diagnostic metadata such as end-effector poses and camera calibration; those are not online observations.

T2 contains 993 curated episodes from a 1000-episode collection. The curation mapping is stored in metadata/curation/grab_roller_993ep.json. The T3 data corresponds to the cp31_symmetric_20260708 collection used by the published epoch-4000 checkpoint.

Repository Layout

data/
├── adjust_bottle/adjust_bottle_200ep/
├── grab_roller/grab_roller_993ep/
├── stack_bowls_two/stack_bowls_two_1000ep/
└── stack_bowls_three/stack_bowls_three_1000ep/
metadata/
├── configs/
└── curation/grab_roller_993ep.json
manifest.json

Each task setting contains data/episode*.hdf5, head-camera preview videos, trajectory caches, scene_info.json, and successful seeds. The top-level layout intentionally matches external/robotwin_local/data/, so it can be copied into a source checkout without renaming task settings. The complete dataset is approximately 78 GB.

Data Fields

Raw episodes include joint actions, three RGB streams, camera intrinsics/extrinsics, cam2world, and left/right end-effector poses. Depending on the collection config, empty diagnostic placeholders may also be present.

Depth, point clouds, end-effector poses, and privileged object state are not ACT observations during evaluation.

Collection and Curation

All episodes passed the task success check. T2 additionally removes seven confirmed abnormal trajectories; prune_manifest.json records removed source IDs and the source-to-release renumbering. T3 and T4 can still contain mild arm interference, contact with a neighboring bowl, wrist-camera occlusion, or variable cuRobo path lengths. These are successful trajectories from the official scene distribution, not guaranteed collision-free demonstrations.

Exact recollection is not guaranteed because successful seeds can depend on SAPIEN, cuRobo, GPU, and driver versions. Use these released demonstrations and the matching source revision when reproducing the published models.

Download and Processing

hf download han-xudong/troncamp-mani-dataset \
  --repo-type dataset \
  --local-dir troncamp-mani-dataset

rsync -a troncamp-mani-dataset/data/ \
  external/robotwin_local/data/

cd external/robotwin_local/policy/ACT
bash process_data.sh adjust_bottle adjust_bottle_200ep 200
bash process_data.sh grab_roller grab_roller_993ep 993
bash process_data.sh stack_bowls_two stack_bowls_two_1000ep 1000
bash process_data.sh stack_bowls_three stack_bowls_three_1000ep 1000

Processing uses the unmodified official ACT loader. The generated processed data is reproducible and is therefore not duplicated in this Dataset repository.

Intended Use

This dataset is intended for reproducing the accompanying ACT checkpoints, studying dual-arm imitation learning, and inspecting geometry-based expert demonstrations in RoboTwin. It is not a real-robot safety dataset and has not been validated for deployment on physical Tron2 hardware.

Versioning and Integrity

manifest.json maps every track to its collection config, episode count, frame statistics, data path, processed setting, companion Model repository, GitHub repository, and release revisions. A release-time checksums.sha256 covers the uploaded files.

License and Attribution

The data contains rendered Tron2 and task assets. The source repository license does not automatically relicense bundled third-party or proprietary assets. Review the GitHub repository LICENSE and NOTICE before redistribution.