Datasets:
TraceSpatial-Trace
RGB images and trajectory question–answer pairs from RoboTracer.
Project · Paper · Code · TraceSpatial-Bench
@@STATUS@@
Overview
TraceSpatial-Trace is the trajectory-focused RGB + QA release associated with RoboTracer. It supports research on predicting object and end-effector motion from images and language. Each displayed row contains an RGB image, a question, and its original answer.
The broader TraceSpatial collection described by the project contains approximately 30 million QA pairs. This repository packages the supplied tracing subset: 3,623,880 QA pairs. It is distinct from the full collection and from TraceSpatial-Bench, the evaluation dataset. Project overview
The paper's data pipeline combines planned object motion in reconstructed CA-1M/ScanNet scenes, traces extracted from DROID/AgiBot manipulation recordings, and RoboTwin simulation. Its three tracing tasks predict image-plane waypoints, predict waypoints with depth, or add depth to supplied 2D waypoints. See the paper's dataset section and tracing appendices for the generation and filtering procedures. Paper
This release contains RGB images and QA text. It does not distribute depth maps, camera calibration, precomputed spatial features, videos, robot actions, or the broader collection's reasoning annotations. Depth values appearing in answers are retained.
Corpus statistics
Counts below come from a complete audit of the five supplied annotation files. “Record” means an original conversation, “QA pair” means one human message and its following assistant answer. Several records can refer to the same image.
| Source | Original records | QA pairs | Unique referenced RGB images |
|---|---|---|---|
| AgiBot | 325,846 | 977,538 | 325,846 |
| CA-1M | 338,419 | 1,015,257 | 20,643 |
| DROID | 105,180 | 105,180 | 20,409 |
| RoboTwin | 146,587 | 1,070,535 | 146,587 |
| ScanNet | 151,790 | 455,370 | 3,730 |
| Total | 1,067,822 | 3,623,880 | 517,215 |
The extracted image directory has 604,660 files. Of these, 87,445 are not referenced by the supplied QA; the conversion only embeds referenced images. All references resolve after adapting 102,228 DROID references from colon-containing timestamps to the underscore filenames present on disk. Original IDs and QA strings are preserved.
Tasks and coordinate conventions
task |
Input in addition to RGB | Answer format |
|---|---|---|
2d_trace |
Task instruction | [(u1, v1), ...] |
3d_trace |
Task instruction | [(u1, v1, d1), ...] |
2d_to_3d |
Task instruction and 2D waypoints | [(u1, v1, d1), ...] |
The paper specifies normalized image-plane coordinates u, v in [0, 1000], with d in meters. The three-component representation is (image coordinate, image coordinate, depth), not Cartesian XYZ. Recovering metric XYZ requires camera intrinsics, which are not included here. The converter preserves values and does not re-normalize, clamp, or reproject them. Trace representation
task is a derived convenience label: the converter checks question wording against answer tuple dimensionality. Ambiguous cases receive null rather than a guessed label. It is not an original annotation field.
Dataset Viewer fields
| Field | Description |
|---|---|
image |
Original encoded RGB image, embedded as a Hugging Face Image feature |
question |
Original human message, unchanged |
answer |
Its following assistant message, unchanged |
type |
Original object or end_effector annotation; null if absent |
source |
agibot, ca1m, droid, robotwin, or scannet |
task |
Derived task label described above |
id |
Original conversation ID; not guaranteed unique |
qa_index |
Zero-based question–answer position within the original conversation |
record_index |
Zero-based record position in its source annotation file |
The tuple (source, record_index, qa_index) uniquely identifies a converted row. Group by (source, record_index) and sort by qa_index to reconstruct the original conversation. The displayed question does not have earlier turns appended; retain the grouped conversation for applications that require dialogue context.
Missing type values are preserved. CA-1M and ScanNet do not provide this field in the supplied files, even though the paper describes their tracing pipeline as object-centric.
| Source | end_effector records |
object records |
Missing type |
|---|---|---|---|
| AgiBot | 274,380 | 51,466 | 0 |
| CA-1M | 0 | 0 | 338,419 |
| DROID | 46,794 | 58,386 | 0 |
| RoboTwin | 107,035 | 39,552 | 0 |
| ScanNet | 0 | 0 | 151,790 |
Loading
from datasets import load_dataset
data = load_dataset(
"leeibo/TraceSpatial-Trace",
"@@CONFIG@@",
split="train",
streaming=True,
)
example = next(iter(data))
print(example["source"], example["type"], example["task"])
print(example["question"])
print(example["answer"])
example["image"].save("example.png")
Use your authenticated HF account while this repository is private. Available configurations are listed in the Viewer/configuration metadata. The conversion scripts produce source-specific configurations and an all configuration for a full conversion; limited runs produce a preview configuration. The initial uploaded preview has 104 QA rows from the first eight records of each source and is illustrative, not randomly sampled.
train is a packaging split, not evidence of an official train/test partition. For evaluation, avoid splitting related images or episodes across partitions. No evaluation results are claimed for this release.
Reproducible conversion and upload
The scripts/ directory contains a multi-process JSON-to-Parquet converter, a resumable concurrent uploader, pinned environment files, and a Chinese usage guide.
The converter streams JSON, preserves original image bytes and QA strings, writes bounded row groups, and checkpoints complete shards. It creates a manifest with per-file SHA-256 hashes, row counts, source statistics, and task-label counts. Original files are never modified. Shards may contain repeated image bytes because each row is independently usable; Parquet compression and dictionary encoding reduce repeated storage within row groups.
Provenance and acknowledgment
This release is associated with the RoboTracer project. We acknowledge the source datasets and assets listed by its official repository: AgiBot World, CA-1M, DROID, RoboTwin 2.0, and ScanNet. The scripts in this repository package the supplied annotations for HF; they do not reproduce the upstream simulation and annotation-generation pipelines.
License
The release-specific dataset license has not yet been supplied by the maintainer. No dataset license is inferred from the project website or code repository. Source data remain subject to their applicable original terms; this card does not grant additional rights.
Citation
Please cite the RoboTracer paper when using this data. The project page uses the earlier title, “RoboTracer: Mastering Spatial Trace with Reasoning in Vision-Language Models for Robotics.” The current arXiv v4 has the updated title below; both refer to arXiv:2512.13660. Version history
@misc{zhou2025robotracer,
title={Towards Spatial Trace with Reasoning in Vision-Language Models for Robotics},
author={Enshen Zhou and Yibo Li and Jingkun An and Jiayuan Zhang and Shanyu Rong and Mengzhen Liu and Yi Han and Yuheng Ji and Huajie Tan and Jiawei He and Pengwei Wang and Zhongyuan Wang and Cheng Chi and Lu Sheng and Shanghang Zhang},
year={2025},
eprint={2512.13660},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2512.13660}
}