artfawl/VintixDatasetII

Vintix II Cross-Domain ICRL Dataset

Dataset Summary

This dataset is a large-scale cross-domain benchmark for in-context reinforcement learning and continuous control. It was introduced with Vintix II and covers a diverse set of tasks spanning robotic manipulation, dexterous control, locomotion, energy management, industrial process control, autonomous driving, and other control settings.

The training set contains 209 tasks across 10 domains, totaling 3.8M episodes and 709.7M timesteps. In addition, the benchmark defines 46 held-out tasks for evaluation on unseen tasks and environment variations.

Domain	Tasks	Episodes	Timesteps	Sample Weight
Industrial-Benchmark	16	288k	72M	10.1%
Bi-DexHands	15	216.2k	31.7M	4.5%
Meta-World	45	670k	67M	9.4%
Kinetix	42	1.1M	62.8M	8.9%
CityLearn	20	146.4k	106.7M	15.0%
ControlGym	9	230k	100M	14.1%
HumEnv	12	120k	36M	5.1%
MuJoCo	11	665.1k	100M	14.1%
SinerGym	22	42.3k	30.9M	4.4%
Meta-Drive	17	271.9k	102.6M	14.4%
Overall	209	3.8M	709.7M	100%

Dataset Structure

The dataset is stored as a collection of .h5 files, where each file corresponds to a single trajectory from a specific environment.

Each trajectory file is split into groups of 10,000 steps, except for the final group, which may contain fewer steps.

Every group contains the following fields:

• proprio_observation: sequence of observations (np.float32)
• action: sequence of actions executed in the environment (np.float32)
• reward: sequence of rewards received after each action (np.float32)
• step_num: step indices within the episode (np.int32)
• demonstrator_action: sequence of demonstrator actions corresponding to the observations

This layout is designed for efficient storage and loading of long trajectories while preserving both collected behavior and demonstrator supervision.