spikefly/SemanticVLA-TraceX-240K-DROID

SemanticVLA TraceX 240K · DROID

🎉 Accepted to [CVPR 2026](https://cvpr.thecvf.com/virtual/2026/poster/39352).<br> ✍️ Fei Ni¹, Zhuo Chen², Yifu Yuan³, Zibin Dong³, Xianze Yao³, Shan Luo², Jianye Hao³, Jiankang Deng¹†, Stefanos Zafeiriou¹†<br> 🏫 ¹Imperial College London &nbsp;&nbsp; ²King's College London &nbsp;&nbsp; ³Tianjin University<br> ✉️ Primary contact: f.ni@imperial.ac.uk

The DROID component of **TraceX-240K** — the trace-annotated trajectory corpus introduced in SemanticVLA. This package is a LeRobot v3.0 repack of DROID · Franka · Open-X-Embodiment DROID v1.0.1 with dense per-frame end-effector trace labels stored directly in every frame row.

📚 Code · checkpoints · paper: https://github.com/Fei-Ni/SemanticVLA_Offcial

🎬 Sample trace overlays

Each clip shows the orange trace label (per-frame gripper position) overlaid on the original episode camera feed.

<table> <tr> <td><img src="assets/traceoverlays/ep005699.gif" width="200" alt="episode 5699"></td> <td><img src="assets/traceoverlays/ep011349.gif" width="200" alt="episode 11349"></td> <td><img src="assets/traceoverlays/ep033979.gif" width="200" alt="episode 33979"></td> <td><img src="assets/traceoverlays/ep039547.gif" width="200" alt="episode 39547"></td> </tr> <tr> <td align="center">_{episode 5699}</td><td align="center">_{episode 11349}</td><td align="center">_{episode 33979}</td><td align="center">_{episode 39547}</td> </tr> </table>

🧭 Trace label schema

Dense per-frame end-effector position columns, stored directly in every frame row:

• —trace.x — float32[1], normalized image-space x coordinate on a [0, 100] scale.
• —trace.y — float32[1], normalized image-space y coordinate on a [0, 100] scale.

(0, 0) is the top-left corner of the frame, (100, 100) the bottom-right.

DROID per-episode camera: unlike Bridge / Fractal / BC-Z (single fixed camera), DROID's trace was projected onto one of the two exterior cameras per episode based on calibration quality. The episode metadata field source.camera_for_trace (exterior_image_1_left or exterior_image_2_left) tells you which camera the trace corresponds to. Overlay accordingly when visualising.

📦 Contents

Video streams

• —observation.images.exterior_image_1_left
• —observation.images.exterior_image_2_left
• —observation.images.wrist_image_left

🧭 Coverage

56,362 episodes with valid dense trace from the local 56K projection selection. All three DROID camera streams (two exterior + wrist) and raw action/observation fields are preserved.

📂 Sibling datasets in TraceX-240K

<table> <tr><th>Dataset</th><th>HF Repository</th><th>Embodiment</th></tr> <tr><td><b>Bridge</b></td><td><a href="https://huggingface.co/datasets/spikefly/SemanticVLA-TraceX-240K-Bridge">spikefly/SemanticVLA-TraceX-240K-Bridge</a></td><td>BridgeData V2 / WidowX</td></tr> <tr><td><b>Fractal</b></td><td><a href="https://huggingface.co/datasets/spikefly/SemanticVLA-TraceX-240K-Fractal">spikefly/SemanticVLA-TraceX-240K-Fractal</a></td><td>Fractal / Google Robot (RT-1)</td></tr> <tr><td><b>BC-Z</b></td><td><a href="https://huggingface.co/datasets/spikefly/SemanticVLA-TraceX-240K-BC-Z">spikefly/SemanticVLA-TraceX-240K-BC-Z</a></td><td>BC-Z / Google Robot</td></tr> </table>

✏️ Citation

If you use this dataset, please cite SemanticVLA:

@inproceedings{ni2026semanticvla,
  title     = {SemanticVLA: Towards Semantic Reasoning over Action Memorization via Synergistic Explicit Trace and Latent Action Planning},
  author    = {Ni, Fei and Chen, Zhuo and Yuan, Yifu and Dong, Zibin and Yao, Xianze and Luo, Shan and Hao, Jianye and Deng, Jiankang and Zafeiriou, Stefanos},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}

📜 License & terms

The trace annotation labels (trace.x / trace.y) are released under the MIT License as part of the SemanticVLA code release. Use of the underlying robot trajectory data is subject to the original source dataset terms (Open-X-Embodiment DROID v1.0.1 (56K dense TraceX selection)). Please consult the upstream dataset license before use.