| --- |
| pretty_name: TriWorldBench Dataset |
| tags: |
| - benchmark |
| - embodied-ai |
| - world-model |
| - robotics |
| - evaluation |
| task_categories: |
| - other |
| --- |
| |
| # TriWorldBench Dataset |
|
|
| Welcome to TriWorldBench, a benchmark for evaluating triple-view embodied world models. |
|
|
| - Paper: [TriWorldBench: A Tri-View Consistency Perspective on Embodied World Models](https://huggingface.co/papers/2609.26314) |
| - Benchmark homepage: [https://www.triworldbench.com](https://www.triworldbench.com) |
| - Benchmark code repository: [TriWorldBench/TriWorldBench](https://github.com/TriWorldBench/TriWorldBench) |
|
|
| ## Overview |
|
|
| TriWorldBench evaluates world models from three synchronized robot views: head, |
| left wrist, and right wrist. The benchmark focuses on whether a model can generate |
| one coherent robot world across all three cameras while preserving task alignment, |
| physical and 3D coherence, motion quality, temporal consistency, and visual quality. |
|
|
| ## Released Files |
|
|
| This Hugging Face dataset provides two compressed dataset bundles: |
|
|
| - `test_dataset`: the official 500-episode test set for leaderboard submission. |
| - `val_dataset`: a 100-episode validation bundle for local development and quick evaluation. |
|
|
| ## `test_dataset` Structure |
| |
| After extraction, the official test dataset is organized as: |
| |
| ```text |
| test_dataset/ |
| +-- data/ |
| | +-- episode1.hdf5 |
| | +-- episode2.hdf5 |
| | +-- ... |
| | +-- episode500.hdf5 |
| +-- first_frame/ |
| | +-- episode1_head.jpg |
| | +-- episode1_left.jpg |
| | +-- episode1_right.jpg |
| | +-- ... |
| | +-- episode500_right.jpg |
| +-- instructions/ |
| +-- episode1.json |
| +-- episode2.json |
| +-- ... |
| +-- episode500.json |
| ``` |
| |
| The test set contains 500 episodes indexed from `episode1` to `episode500`. |
| For each episode, `data/episodeN.hdf5` contains the trajectory/action sequence, |
| `first_frame/` provides the initial 320x240 observations for the head, left wrist, |
| and right wrist cameras, and `instructions/episodeN.json` contains the task instruction. |
|
|
| Participants should run their models on this test set and submit generated videos |
| for all three views: `head.mp4`, `left.mp4`, and `right.mp4`. |
|
|
| ## `val_dataset` Structure |
| |
| The validation bundle contains 100 episodes derived from RoboTwin2.0, together |
| with the corresponding ground-truth data and processed state annotations. |
| After extraction, it is organized as: |
| |
| ```text |
| val_dataset/ |
| +-- test_dataset/ |
| | +-- data/ |
| | +-- first_frame/ |
| | +-- instructions/ |
| +-- STATE/ |
| | +-- episode*.json |
| +-- gt_dataset/ |
| +-- episode*/ |
| ``` |
| |
| The `val_dataset/test_dataset` folder provides the same type of model inputs as |
| the official test set. The bundled `gt_dataset` and `STATE` folders are already |
| formatted for the TriWorldBench evaluation code. Matching VQA question annotations |
| for this validation set are included in the benchmark code repository under |
| `metrics/VQA/qa_val/`. |
| |
| ## Evaluation and Submission |
| |
| Please refer to the benchmark homepage and code repository for environment setup, |
| inference, evaluation, and leaderboard submission instructions. |
