REKI-2B RoboTwin Clean+Rand

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 the clean and randomized RoboTwin 2.0 demonstrations (dual-arm AgileX ALOHA, simulation, 50 tasks). We evaluate it on clean scenes (Easy) and randomized scenes (Hard).

For the out-of-distribution setting (finetuned on clean scenes only), use kyutai/reki-2b-robotwin-clean. 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	AgileX ALOHA, two 6-DoF arms with parallel grippers
Cameras	head camera, two wrist cameras
State (14), `observation.state`	absolute joint positions `[left arm (6), left gripper (1), right arm (6), right gripper (1)]`
Action (14)	absolute joint positions `[left arm (6), left gripper (1), right arm (6), right gripper (1)]`
Control	15 Hz, 15 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 robotwin_cleanrand entry of launchers/sft.py hold the hyperparameters.

Setting	Value
Dataset	RoboTwin 2.0 demonstrations: 50 clean and 500 randomized per task
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	24 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-robotwin-cleanrand --local-dir reki-2b-robotwin-cleanrand
pixi run python launchers/eval_protocols.py robotwin reki-2b-robotwin-cleanrand 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 RoboTwin 2.0, 50 tasks, 20 episodes per task, unseen instructions.

Success rate (%)	Easy	Hard
REKI	85.9	84.6

Per-task results

Per-task success rates

Success rate per task, 20 episodes per task, unseen instructions.

Success rate (%)	Easy	Hard
`adjust_bottle`	100.0	100.0
`beat_block_hammer`	95.0	95.0
`blocks_ranking_rgb`	95.0	100.0
`blocks_ranking_size`	95.0	90.0
`click_alarmclock`	100.0	100.0
`click_bell`	100.0	100.0
`dump_bin_bigbin`	60.0	75.0
`grab_roller`	100.0	95.0
`handover_block`	70.0	35.0
`handover_mic`	100.0	95.0
`hanging_mug`	30.0	25.0
`lift_pot`	100.0	95.0
`move_can_pot`	40.0	40.0
`move_pillbottle_pad`	90.0	85.0
`move_playingcard_away`	100.0	100.0
`move_stapler_pad`	75.0	65.0
`open_laptop`	100.0	100.0
`open_microwave`	95.0	95.0
`pick_diverse_bottles`	55.0	85.0
`pick_dual_bottles`	70.0	65.0
`place_a2b_left`	90.0	100.0
`place_a2b_right`	90.0	100.0
`place_bread_basket`	85.0	90.0
`place_bread_skillet`	100.0	95.0
`place_burger_fries`	100.0	100.0
`place_can_basket`	80.0	70.0
`place_cans_plasticbox`	100.0	100.0
`place_container_plate`	100.0	100.0
`place_dual_shoes`	95.0	70.0
`place_empty_cup`	100.0	100.0
`place_fan`	95.0	100.0
`place_mouse_pad`	50.0	45.0
`place_object_basket`	75.0	60.0
`place_object_scale`	100.0	80.0
`place_object_stand`	100.0	95.0
`place_phone_stand`	95.0	100.0
`place_shoe`	100.0	100.0
`press_stapler`	95.0	95.0
`put_bottles_dustbin`	55.0	40.0
`put_object_cabinet`	75.0	75.0
`rotate_qrcode`	90.0	100.0
`scan_object`	40.0	85.0
`shake_bottle`	100.0	100.0
`shake_bottle_horizontally`	100.0	100.0
`stack_blocks_three`	100.0	100.0
`stack_blocks_two`	100.0	100.0
`stack_bowls_three`	90.0	75.0
`stack_bowls_two`	100.0	95.0
`stamp_seal`	75.0	80.0
`turn_switch`	50.0	40.0
Mean	85.9	84.6

Limitations

We evaluate 20 trials per task. The RoboTwin leaderboard uses 100.
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}
}