suryatmodulus/reki-2b-robocasa365

REKI-2B RoboCasa365

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 RoboCasa365 (Franka Panda arm on an Omron mobile base, simulated kitchens). We evaluate it on the 50 target tasks of the official multitask protocol.

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	PandaOmron, Franka Panda arm on an Omron mobile base
Cameras	left and right agent views, wrist view
State (16), observation.state	base pose (position 3 + quaternion 4), EE pose in base frame (position 3 + quaternion 4), gripper finger positions (2)
Action (12)	base motion (4), control-mode flag (1), delta EE position (3), delta EE rotation (3), gripper (1)
Control	20 Hz, 20 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 robocasa365 entry of launchers/sft.py hold the hyperparameters.

Setting	Value
Dataset	RoboCasa365 human demonstrations of the 300 pretraining tasks (65 atomic, 235 composite; 32,043 episodes)
Steps	25,000
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, left-right mirroring (details)
GPUs	32 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-robocasa365 --local-dir reki-2b-robocasa365
pixi run python launchers/eval_protocols.py robocasa365 reki-2b-robocasa365 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 50 target tasks of RoboCasa365 (18 atomic-seen, 16 composite-seen, 16 composite-unseen) in the pretraining kitchens, 50 episodes per task, official per-task horizons. Mean is the mean over the 50 tasks.

Success rate (%)	Atomic-seen	Composite-seen	Composite-unseen	Mean
REKI	65.8	32.0	14.1	38.4

Per-task results

Per-task success rates

Success rate per task, 50 episodes per task.

Group	Task	Success rate (%)
Atomic-seen	CloseBlenderLid	28.0
Atomic-seen	CloseFridge	88.0
Atomic-seen	CloseToasterOvenDoor	66.0
Atomic-seen	CoffeeSetupMug	42.0
Atomic-seen	NavigateKitchen	22.0
Atomic-seen	OpenCabinet	86.0
Atomic-seen	OpenDrawer	52.0
Atomic-seen	OpenStandMixerHead	98.0
Atomic-seen	PickPlaceCounterToCabinet	82.0
Atomic-seen	PickPlaceCounterToStove	68.0
Atomic-seen	PickPlaceDrawerToCounter	68.0
Atomic-seen	PickPlaceSinkToCounter	90.0
Atomic-seen	PickPlaceToasterToCounter	88.0
Atomic-seen	SlideDishwasherRack	82.0
Atomic-seen	TurnOffStove	20.0
Atomic-seen	TurnOnElectricKettle	78.0
Atomic-seen	TurnOnMicrowave	58.0
Atomic-seen	TurnOnSinkFaucet	68.0
Composite-seen	DeliverStraw	4.0
Composite-seen	GetToastedBread	8.0
Composite-seen	KettleBoiling	20.0
Composite-seen	LoadDishwasher	48.0
Composite-seen	PackIdenticalLunches	20.0
Composite-seen	PreSoakPan	48.0
Composite-seen	PrepareCoffee	10.0
Composite-seen	RinseSinkBasin	46.0
Composite-seen	ScrubCuttingBoard	38.0
Composite-seen	SearingMeat	2.0
Composite-seen	SetUpCuttingStation	54.0
Composite-seen	StackBowlsCabinet	66.0
Composite-seen	SteamInMicrowave	18.0
Composite-seen	StirVegetables	30.0
Composite-seen	StoreLeftoversInBowl	38.0
Composite-seen	WashLettuce	62.0
Composite-unseen	ArrangeBreadBasket	12.0
Composite-unseen	ArrangeTea	20.0
Composite-unseen	BreadSelection	32.0
Composite-unseen	CategorizeCondiments	2.0
Composite-unseen	CuttingToolSelection	4.0
Composite-unseen	GarnishPancake	12.0
Composite-unseen	GatherTableware	2.0
Composite-unseen	HeatKebabSandwich	0.0
Composite-unseen	MakeIceLemonade	32.0
Composite-unseen	PanTransfer	0.0
Composite-unseen	PortionHotDogs	4.0
Composite-unseen	RecycleBottlesByType	20.0
Composite-unseen	SeparateFreezerRack	0.0
Composite-unseen	WaffleReheat	20.0
Composite-unseen	WashFruitColander	42.0
Composite-unseen	WeighIngredients	24.0

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}
}