REKI: Registering Knowledge in VLAs

A 2.4B vision-language-action model pretrained on public data only, with a memory of its recent past at negligible cost.

Paper · Project page · Code · Models

Installation

Requires Linux x86-64, Python 3.12, an NVIDIA GPU of the Ampere generation or newer, a C compiler and the Python 3.12 headers. The default PyPI build of PyTorch needs NVIDIA driver 580 or newer (CUDA 13). With driver 570, install a CUDA 12.8 build first: pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cu128.

pip install reki-vla    # or: uv add reki-vla

Quickstart

from reki import RekiPolicy

policy = RekiPolicy.from_pretrained("kyutai/reki-2b-libero", device="cuda")
policy.reset()                    # at the start of every episode: clears the register memory
actions = policy.predict(
    images={"camera_top": img0, "camera_wrist_left": img1},  # HxWx3 uint8, RGB
    state=proprio,                # training dataset's observation.state, float [policy.spec.state_dim]
    task="put the bowl on the plate",
)                                 # float32 np.ndarray [policy.n_action_steps, action_dim], unnormalized

• Call predict once every policy.n_action_steps environment steps (1 s for all released models), and execute the returned actions in between.
• policy.spec.camera_keys and policy.spec.state_layout give the expected cameras and the state order. The images must have the orientation of the training data (for LIBERO, rotate the env images by 180°).

Evaluation, training and real-robot deployment use the repository: see the README.

License

• Code: MIT. Two files and parts of one file come from LeRobot, and parts of three files from Transformers, under Apache-2.0 (NOTICE and LICENSE-APACHE in the package).
• Weights: CC-BY-NC-SA 4.0. The weights derive from Qwen3.5-2B and third-party datasets. Each keeps its own license.