KlaskLab/klasktron-il-benchmark

KlaskTron

KlaskTron is a Klask imitation-learning dataset with three tiers: real human play, axis-mirror-augmented human data, and large-scale synthetic rollouts.

Tier	Episodes	Steps	Source
human	111	80,591	reconstructed from recorded human play
human_augmented	444	322,364	axis-mirror augmentation of the human tier (none, x, y, xy)
synthetic	34,649	5,000,000	generated in simulation from an expert-like policy

Total: 35,204 episodes and 5,402,955 action steps.

Loading Parquet

from datasets import load_dataset

human = load_dataset("KlaskLab/klasktron-il-benchmark", "human", split="train")
human_augmented = load_dataset("KlaskLab/klasktron-il-benchmark", "human_augmented", split="train")
synthetic = load_dataset("KlaskLab/klasktron-il-benchmark", "synthetic", split="train")

Each Parquet row is one transition:

Column	Meaning
tier	human, human_augmented, or synthetic
episode_id	tier-local episode id
source_file	original pickle file
episode_in_file	episode index inside that pickle
step	transition index inside the episode
obs	state at time t, shape [12]
act	action at time t, shape [4]
next_obs	state at time t + 1, shape [12]
rew	all-zero placeholder
terminal	true only on the final transition of a terminal episode

Observation layout:

[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:

[peg1_vx, peg1_vy, peg2_vx, peg2_vy]

Loading Pickles

Raw pickle files live under:

human/trajectories/
human_augmented/trajectories/
synthetic/trajectories/

Episode fields:

• obs: float32 array, shape [T + 1, 12]
• acts: float32 array, shape [T, 4]
• rews: float32 array, shape [T] where present
• infos: optional per-step metadata
• terminal: whether the rollout ends naturally

Raw container types:

Tier	Container
human	imitation.data.types.TrajectoryWithRew
human_augmented	imitation.data.types.Trajectory
synthetic	plain Python dict

The human and human_augmented pickles need a compatible Python environment with imitation available. The synthetic pickles only use plain dictionaries and NumPy arrays. As usual, only unpickle data you trust.

Rewards

There is no meaningful per-step reward signal. rew / rews is an all-zero placeholder and should not be used for offline-RL returns, filtering, or scoring.

Notes

• The human tier comes from 111 recorded human Klask trajectories that were reestimated into simulation coordinates.
• The human_augmented tier extends the human tier 4x by applying axis-mirror augmentation (none, x, y, xy) to each trajectory.
• The synthetic tier reflects simulator and expert-policy assumptions.

Why Both Formats Exist

The pickle files are the original episode objects used by imitation-learning code that expects imitation.data.types.Trajectory / TrajectoryWithRew objects. The mirrored Parquet folder exposes the same trajectories as a safer, easier-to-inspect transition table for Hugging Face Datasets, pandas, Polars, DuckDB, PyArrow, browser previews, and Croissant-compatible tooling.

License

Creative Commons Attribution 4.0 International License (CC-BY-4.0).

Citation

@article{klasktron2026,
  title={KlaskTron: A Contact-Rich, Adversarial Benchmark for Imitation Learning},
  author={Anonymous},
  year={2026},
  note={Under review at NeurIPS 2026}
}