KyutaiREKI
REKI 2B RoboTwin Clean + Random
This dual-arm policy trains on both clean and randomized RoboTwin scenes. Randomized evaluation is therefore not the same unseen-scene test as for the clean-only model.
Research release
Register tokens carry recent visual context
REKI stores 32 tokens per frame across three frames one second apart. The action expert can use this recent history. The repository separates raw and moving-average weights; the documented fine-tuning path starts with raw weights.reki-2b-robotwin-cleanrand: Kyutai model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | kyutai/reki-2b-robotwin-cleanrand |
| Developers | Kyutai, Inria and École nationale des ponts et chaussées.reki-2b-robotwin-cleanrand: Kyutai model card |
| Actual size | 2.4B parameters: a 2.2B vision-language component and 0.2B action expert.reki-2b-robotwin-cleanrand: Kyutai model card |
| Ancestry and license | Qwen 3.5 2B; CC BY-NC-SA 4.0 weights and MIT code.reki-2b-robotwin-cleanrand: Kyutai model card |
Related models
A Qwen-based vision-language-action model pretrained on robot, simulation, human-hand and vision-language data.
Launch date not confirmedResearch releaseDownloadable