REKI gives robot actions a short visual memory
Kyutai and its research partners introduced REKI, a robot-control model with 2.4 billion parameters.
In this article
How the memory works
Introduced in October 2026, REKI translates camera views and a written instruction into sequences of robot actions. The research paper describes register tokens: compact summaries of observations, stored inside the component that predicts movements. Later action sequences can read this retained context without processing the old camera frames again through the model that interprets images and language.
Adapt the model to the robot
The separate training manual describes adapting REKI to another robot: convert and register demonstration data in LeRobot v3.0 format, configure the dataset and robot state layout, train, then serve the trained model. An existing state layout must retain its dimension; a new layout gets a newly trained input network.
Exporting with the dataset or test environment specified also includes settings for running the model and statistics used to scale its data. The exported model can then be used by RekiPolicy without retaining the training data.
What running it requires
The implementation README requires Linux x86-64, an NVIDIA GPU with CUDA 12, and pixi to manage the software environment. Its example resets the memory at each episode’s start before predicting actions from images, robot state and the task. Camera names, state order and image orientation must match the model’s training setup.
One release limitation matters: the repository says the pretraining code and data pipeline will arrive in a later update.
Sources(5)
- Robot-memory coverageChecked:
- Robot-memory overviewChecked:
- Robot-memory researchChecked:
- Implementation notesChecked:
- Adaptation guideChecked:
