spikefly/SemanticVLA-TraceX-260K

SemanticVLA — TraceX 260K

🚧 Placeholder repository. The URL is stable and safe to embed in posters / slides; data will be uploaded incrementally per the release roadmap.

SemanticVLA-TraceX-260K aggregates four large-scale manipulation datasets, converted to the LeRobot format and augmented with dense end-effector trace annotations that align to every (episode, step) index. It is the supervision signal used to train the trace head and the trace-conditioned Latent Action Model (LAM) in SemanticVLA.

This repository hosts the all-in-one bundle. If you only need a single subset, prefer the dedicated per-dataset repos — they share the same schema and are bit-identical to the corresponding subdirectories here:

• 🤗 SemanticVLA-Bridge-LeRobot — BridgeData V2, WidowX, ~53k episodes
• 🤗 SemanticVLA-Fractal-LeRobot — RT-1 / Fractal, Google Robot, ~87k episodes
• 🤗 SemanticVLA-BCZ-LeRobot — BC-Z, 39,350 episodes
• 🤗 SemanticVLA-DROID-LeRobot — DROID, Franka Panda, partial release first

Other resources:

• Code: https://github.com/Fei-Ni/SemanticVLA_Offcial
• Checkpoints: https://hf.co/collections/spikefly/semanticvla-model-zoo
• Collection · Datasets: https://hf.co/collections/spikefly/semanticvla-datasets
• Collection · Model Zoo: https://hf.co/collections/spikefly/semanticvla-model-zoo

Sub-datasets

Subset	Robot	Episodes	Frames	Source
bridge_orig	WidowX	~53,000	~2.0M	BridgeData V2 (OXE bridge_orig)
fractal20220817	Google Robot	~87,000	~3.8M	RT-1 (OXE fractal20220817_data)
bc_z	xArm	~39,000	~5.5M	BC-Z (OXE bc_z)
droid	Franka	~92,000	~28M	DROID (partial release first)

The exact subset sizes will be finalized in the dataset card upon upload.

Planned layout

SemanticVLA-TraceX-260K/
├── bridge_orig/                # LeRobot dataset
│   ├── meta/
│   ├── data/
│   └── videos/
├── fractal20220817/            # LeRobot dataset
├── bc_z/                       # LeRobot dataset
├── droid/                      # LeRobot dataset (partial first)
├── traces/                     # Per-dataset trace NPY index
│   ├── bridge_coords.npy
│   ├── bridge_offsets.npy
│   ├── bridge_present.npy
│   ├── fractal_coords.npy
│   ├── fractal_offsets.npy
│   ├── fractal_present.npy
│   ├── bcz_coords.npy
│   ├── bcz_offsets.npy
│   └── bcz_present.npy
└── README.md

Every LeRobot sub-dataset follows the standard info.json / episodes.jsonl / tasks.jsonl / stats.json / modality.json schema. The shared trace index uses memory-mapped NPY arrays:

• <name>_coords.npy — (T, 2) flat array of trace (x, y) coordinates in image space.
• <name>_offsets.npy — per-episode offset into coords so a single (episode, step) lookup is O(1).
• <name>_present.npy — boolean mask of which episodes have a trace.

A minimal Python loader will be shipped with the code repo.

How to use (placeholder API)

from huggingface_hub import snapshot_download

local_root = snapshot_download(
    repo_id="spikefly/SemanticVLA-TraceX-260K",
    repo_type="dataset",
    allow_patterns=["bridge_orig/*", "traces/bridge_*"],
)
# then load with `lerobot.common.datasets.lerobot_dataset.LeRobotDataset`

Citation

@misc{semanticvla2026,
  title  = {SemanticVLA: Semantic Vision-Language-Action Models with Trace-Conditioned Latent Action Tokens},
  author = {SemanticVLA Authors},
  year   = {2026},
  url    = {https://github.com/Fei-Ni/SemanticVLA_Offcial}
}

Please also cite the original dataset papers (BridgeData V2, RT-1 / Fractal, BC-Z, DROID, Open X-Embodiment) when using individual subsets.

License

Released under the MIT License, subject to the upstream licenses of each constituent dataset.