REKI is a 2.4-billion-parameter vision-language-action model (VLA) that the authors say sets a new Pareto frontier between performance and model size. It was pretrained exclusively on public data, drawing on a curated mix of robot demonstrations, egocentric human video, and vision-language data. On the LIBERO-Plus, RoboTwin Clean2Rand, and RoboCasa365 benchmarks, REKI is outperformed only by models roughly twice its size.
The architecture pairs a vision-language model with a flow-matching action expert. The authors add a lightweight memory: a few register tokens in the action expert summarize the VLM's output at each frame and persist in the expert's cache. They report that this memory improves success rates in both simulation and on real robots, and makes oscillations between poses about four times rarer.
The design is intentionally simple and optimized for training and inference efficiency. The authors release the full code for data curation, training, inference, evaluation, and deployment, along with all pretrained and finetuned weights.