REKI-2B LIBERO

REKI is a 2.4B vision-language-action model with register tokens in its action expert, pretrained on public data only. This model is kyutai/reki-2b finetuned on LIBERO (Franka Panda, simulation). One model covers the four suites: Spatial, Object, Goal and Long. We evaluate it on LIBERO and on the seven perturbation axes of LIBERO-Plus.

The kyutai/reki-2b card describes the architecture and the pretraining.

Model Details

Developed by: Kyutai, Inria, École nationale des ponts et chaussées
Model type: Vision-language-action model (flow matching)
Parameters: 2.4B
Finetuned from: kyutai/reki-2b
Memory window: K = 3 frames, 1 s apart
License: CC BY-NC-SA 4.0 (weights), MIT (code)

Model Sources

Code: github.com/kyutai-labs/reki
Paper: REKI: Registering Knowledge in VLAs
Project page: kyutai.org/reki

Embodiment

Item	Value
Robot	Franka Panda, single arm
Cameras	agent view, wrist view
State (8), `observation.state`	EE position (3), EE axis-angle rotation (3), gripper joint positions (2)
Action (7)	delta EE position (3), delta axis-angle rotation (3), gripper (1)
Control	10 Hz, 10 steps per chunk

Files

model.safetensors (EMA merged) and train_config.yaml.
reki_inference.json and reki_stats.safetensors: the inference spec and the normalization stats for RekiPolicy.
config.json, policy_preprocessor.json and policy_postprocessor.json: the files for the LeRobot plugin.

Training

configs/sft.yaml and the libero entry of launchers/sft.py hold the hyperparameters.

Setting	Value
Dataset	`HuggingFaceVLA/libero`
Steps	15,000
Learning rate	1e-4 (action expert), 3e-5 (backbone)
Schedule	warmup-stable-decay: linear warmup over 1,000 steps, cosine decay over the last 5,000
Augmentation	color jitter, camera dropout, prompt and state dropout, random crop and scale (details)
GPUs	8 H100

Evaluation

Install the repository (README) and the simulator (scripts/eval/SETUP.md). Then set the simulator variables of the evaluation docs.

pixi run hf download kyutai/reki-2b-libero --local-dir reki-2b-libero
pixi run python launchers/eval_protocols.py libero reki-2b-libero evals | bash
pixi run python launchers/eval_protocols.py libero_plus reki-2b-libero evals | bash
pixi run python scripts/eval/results_table.py evals

eval_protocols.py prints the rollout.py commands of the released protocol, one GPU each.

Results

Success rate on the 4 LIBERO suites, 10 tasks per suite, 50 episodes per task.

Success rate (%)	Spatial	Object	Goal	Long	Mean
REKI	99.2	99.4	96.6	95.0	97.6

Success rate on LIBERO-Plus, 4 suites × 7 perturbation axes, 10,030 perturbed episodes in total. Overall counts all episodes.

Success rate (%)	Camera	Robot init	Language	Light	Background	Noise	Layout	Overall
REKI	93.2	75.6	88.5	97.7	98.4	98.6	84.7	90.4

Limitations

We evaluate the model in simulation only. There is no real-robot or safety evaluation.

Citation

@article{marrie2026reki,
  title   = {{REKI}: Registering Knowledge in {VLAs}},
  author  = {Marrie, Juliette and Pacaud, Paul and Ramanana Rahary, Adrien and Chen, Shizhe and B{\"o}hle, Moritz},
  journal = {arXiv:2610.22178},
  year    = {2026}
}