Robot learning dataset
1X World Model Compression Challenge Dataset
Real-world humanoid interaction recordings distributed as tokenized visual observations and robot-state arrays for the 1X World Model Compression Challenge. The current v2.0 training and validation release uses NVIDIA Cosmos tokens; older v1.1 data used MAGVIT2. The tokenized dataset is Apache-2.0, while the separate raw-video release is non-commercial.
Records last verified Oct 13, 2026 · Send a correction
Licence & use
Apache-2.0
Commercial use allowed
Licence text ↗Terms apply to the checked release. Read publisher conditions for mixed collections, converted formats and gated access.
- Publisher
- 1X Technologies
- Released
- Not dated
- Published size
- See size notes
- Modalities
- rgb, proprioception
- Formats
- Binary arrays, JSON
- Real / simulated
- Real
- Includes humanoid recordings
- Yes
Size & release notes
Current card publishes ~100 hours for train/val v2.0, 450 test samples and 27.3 GB total repository files. No exact episode total, duration or TB conversion is stored. The original challenge repository describes EVE, but the v2.0 card does not identify the current robot model; no catalog robot slug is assigned.
Recorded embodiments
- 1X humanoid (current v2.0 model unspecified)
More robot learning datasets
Evidence
Sources
- [s1] 1X World Model Compression Challenge tokenized dataset card ↗
1X Technologies · official · accessed Oct 13, 2026
- [s2] 1X World Model Challenge official repository ↗
1X Technologies · official · accessed Oct 13, 2026
- [s3] Apache License, Version 2.0 ↗
Apache Software Foundation · official · accessed Oct 13, 2026