Update README.md

README.md

@@ -1,3 +1,166 @@
----
-license: cc-by-4.0
----
+---
+license: cc-by-4.0
+pretty_name: KlaskTron
+source_datasets:
+  - original
+size_categories:
+  - 1M<n<10M
+tags:
+  - imitation-learning
+  - reinforcement-learning
+  - robotics
+  - control
+  - simulation
+  - klask
+  - pickle
+task_categories:
+  - robotics
+  - reinforcement-learning
+---
+
+# KlaskTron
+
+## Dataset Summary
+
+KlaskTron is a trajectory dataset for imitation-learning experiments in the Klask tabletop game. It includes three tiers of demonstrations stored in simulation coordinates so users can compare learning from real play, augmented real data, and large-scale synthetic rollouts.
+
+| Tier | Files | Episodes | Steps | Source |
+| --- | ---: | ---: | ---: | --- |
+| `human` | 106 | 106 | 81,455 | reconstructed from recorded human play |
+| `augmented` | 7 | 424 | 325,820 | augmented from the human tier |
+| `synthetic` | 542 | 34,649 | 5,000,000 | generated in simulation from an expert or expert-like policy |
+
+Across all tiers, the repository contains 35,179 episodes and 5,407,275 action steps.
+
+## Repository Structure
+
+```text
+klasktron/
+|-- human/
+|   |-- trajectories/
+|   `-- metadata.json
+|-- augmented/
+|   |-- trajectories/
+|   `-- metadata.json
+|-- synthetic/
+|   |-- trajectories/
+|   `-- metadata.json
+|-- dataset_card.md
+`-- README.md
+```
+
+Each tier contains a `trajectories/` directory with `.pkl` files and a `metadata.json` summary file. The `human` tier uses per-episode files such as `reestimate_*.pkl`, while the `augmented` and `synthetic` tiers are stored as chunked files such as `chunk_00000.pkl`.
+
+## Data Format
+
+Trajectory files are stored as Python pickle files. A record corresponds to
+one episode and carries the following fields:
+
+- `obs`: float32 array with shape `[T + 1, 12]`
+- `acts`: float32 array with shape `[T, 4]`
+- `rews`: float32 array with shape `[T]` (zero placeholder, see Rewards below)
+- `infos`: optional per-step metadata (null in every released episode)
+- `terminal`: whether the rollout ends naturally
+
+Observation layout (identical across all three tiers):
+
+```text
+[ball_x, ball_y,
+ ball_vx, ball_vy,
+ peg1_x, peg1_y,
+ peg2_x, peg2_y,
+ peg1_vx, peg1_vy,
+ peg2_vx, peg2_vy]
+```
+
+Action layout (identical across all three tiers):
+
+```text
+[peg1_vx, peg1_vy, peg2_vx, peg2_vy]
+```
+
+Per-tier container type:
+
+| Tier | Container type |
+| --- | --- |
+| `human` | `imitation.data.types.TrajectoryWithRew` |
+| `augmented` | `imitation.data.types.Trajectory` |
+| `synthetic` | plain Python `dict` |
+
+### Rewards
+
+The `rews` field is an all-zero placeholder in every tier that stores it
+(the `augmented` tier uses `Trajectory` and therefore does not store a
+`rews` field at all). No tier carries a meaningful per-step reward signal,
+and consumers should not use `rews` as returns for offline RL, filtering
+by return, or scoring.
+
+### Loading notes
+
+Human- and augmented-tier files reference `imitation.data.types.*` symbols
+at unpickling time, so they need a Python environment with the `imitation`
+package (plus `numpy`) installed. Synthetic-tier files use only plain dicts
+and ndarrays and can be loaded with the standard-library `pickle` alone.
+
+## Data Collection and Processing
+
+### Human Tier
+
+- Recorded from real human Klask play
+- Converted into simulation-coordinate trajectories through a separate state-reestimation pipeline
+- Manually cleaned before inclusion in this release
+
+### Augmented Tier
+
+- Derived from the human tier
+- Produced through augmentation and benchmark preprocessing
+- Intended to extend the small real-data regime without changing the task format
+
+### Synthetic Tier
+
+- Generated in simulation from an expert or expert-like policy
+- Metadata identifies the generator as `rl_games_expert` on `Isaac-Klask-v0`
+- Generated with self-play enabled
+
+## Recommended Uses
+
+- Benchmarking behavioral cloning and related imitation-learning methods
+- Comparing human, augmented, and synthetic demonstrations under one shared format
+- Studying data-scaling behavior and sim-to-sim transfer within the Klask benchmark
+
+## Out-of-Scope Uses
+
+- Real-world deployment or safety-critical control
+- Human identity inference or user profiling
+- Claims about human skill outside this benchmark and preprocessing pipeline
+
+## Limitations
+
+- The human tier depends on a separate state-reestimation and cleaning pipeline.
+- The augmented tier inherits the biases and coverage limits of the human data.
+- The synthetic tier reflects simulator assumptions and expert-policy biases.
+- Because the data is stored as pickle files, it is better suited for download-and-load workflows than direct browser inspection.
+- None of the three tiers carry a meaningful per-step reward signal. The `rews` field is an all-zero placeholder wherever it appears (`human` and `synthetic`); the `augmented` tier does not store rewards at all.
+
+## Related Resources
+
+- `klask_imitation`: main codebase for training, evaluation, augmentation, 
+  and synthetic data generation. URL: [Anonymized for review]
+- `klask_reestimate`: reestimation pipeline used to produce the human 
+  trajectories from recorded gameplay. URL: [Anonymized for review]
+  
+## Licensing Information
+
+This dataset is released under the Creative Commons Attribution 4.0 
+International License (CC-BY-4.0).
+
+## Citation
+
+```bibtex
+@article{klasktron2026,
+  title={KlaskTron: A Contact-Rich, Adversarial Benchmark for Imitation Learning},
+  author={Anonymous},
+  year={2026},
+  note={Under review at NeurIPS 2026}
+}
+```