REKI-2B RoboTwin Clean

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 RoboTwin 2.0 demonstrations only (dual-arm AgileX ALOHA, simulation, 50 tasks). We evaluate it on clean scenes (Easy), on the four single-axis randomizations of RoboTwin Clean2Rand, and on fully randomized scenes (Hard).

For in-distribution results on randomized scenes, use kyutai/reki-2b-robotwin-cleanrand. 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_clean entry of launchers/sft.py hold the hyperparameters.

Setting	Value
Dataset	RoboTwin 2.0 clean demonstrations, 50 per task
Steps	20,000
Learning rate	1e-4 (action expert), 2e-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	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-robotwin-clean --local-dir reki-2b-robotwin-clean
pixi run python launchers/eval_protocols.py robotwin reki-2b-robotwin-clean evals | bash
pixi run python launchers/eval_protocols.py robotwin_clean2rand reki-2b-robotwin-clean 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 and RoboTwin Clean2Rand, 50 tasks, 20 episodes per task, unseen instructions. Clean2Rand applies one randomization (background, light, clutter, table height) to a clean scene. Hard applies all of them.

Success rate (%)	Easy	Background	Light	Clutter	Height	Hard	Mean
REKI	73.4	71.3	73.5	57.8	69.9	48.7	65.8

Per-task results

Per-task success rates

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

Success rate (%)	Easy	Background	Light	Clutter	Height	Hard
`adjust_bottle`	100.0	100.0	100.0	100.0	100.0	100.0
`beat_block_hammer`	80.0	95.0	80.0	70.0	60.0	40.0
`blocks_ranking_rgb`	75.0	90.0	100.0	90.0	95.0	80.0
`blocks_ranking_size`	80.0	75.0	70.0	20.0	75.0	5.0
`click_alarmclock`	100.0	100.0	100.0	90.0	100.0	70.0
`click_bell`	100.0	100.0	100.0	95.0	90.0	65.0
`dump_bin_bigbin`	55.0	85.0	70.0	25.0	35.0	15.0
`grab_roller`	95.0	100.0	100.0	75.0	95.0	90.0
`handover_block`	55.0	75.0	55.0	10.0	55.0	5.0
`handover_mic`	100.0	95.0	100.0	60.0	100.0	35.0
`hanging_mug`	20.0	10.0	25.0	20.0	15.0	10.0
`lift_pot`	100.0	85.0	80.0	65.0	65.0	30.0
`move_can_pot`	30.0	15.0	25.0	35.0	30.0	30.0
`move_pillbottle_pad`	85.0	50.0	70.0	30.0	65.0	5.0
`move_playingcard_away`	100.0	95.0	95.0	80.0	95.0	70.0
`move_stapler_pad`	25.0	35.0	50.0	35.0	50.0	20.0
`open_laptop`	100.0	90.0	85.0	95.0	85.0	85.0
`open_microwave`	60.0	35.0	50.0	25.0	40.0	65.0
`pick_diverse_bottles`	55.0	45.0	35.0	50.0	75.0	50.0
`pick_dual_bottles`	75.0	35.0	40.0	55.0	85.0	55.0
`place_a2b_left`	85.0	90.0	90.0	60.0	75.0	55.0
`place_a2b_right`	85.0	95.0	85.0	60.0	80.0	60.0
`place_bread_basket`	80.0	85.0	80.0	80.0	80.0	50.0
`place_bread_skillet`	80.0	85.0	85.0	75.0	85.0	55.0
`place_burger_fries`	95.0	100.0	95.0	90.0	95.0	90.0
`place_can_basket`	55.0	50.0	65.0	10.0	70.0	5.0
`place_cans_plasticbox`	100.0	80.0	100.0	50.0	100.0	20.0
`place_container_plate`	90.0	100.0	95.0	85.0	80.0	60.0
`place_dual_shoes`	25.0	35.0	40.0	15.0	30.0	10.0
`place_empty_cup`	100.0	95.0	90.0	80.0	90.0	55.0
`place_fan`	65.0	45.0	65.0	45.0	65.0	55.0
`place_mouse_pad`	35.0	45.0	45.0	30.0	35.0	30.0
`place_object_basket`	80.0	90.0	80.0	55.0	50.0	50.0
`place_object_scale`	90.0	70.0	85.0	50.0	75.0	30.0
`place_object_stand`	90.0	90.0	85.0	85.0	80.0	65.0
`place_phone_stand`	90.0	75.0	85.0	65.0	90.0	50.0
`place_shoe`	45.0	50.0	35.0	60.0	40.0	55.0
`press_stapler`	85.0	75.0	90.0	100.0	70.0	65.0
`put_bottles_dustbin`	45.0	30.0	35.0	40.0	25.0	10.0
`put_object_cabinet`	35.0	25.0	35.0	15.0	30.0	45.0
`rotate_qrcode`	75.0	85.0	80.0	55.0	95.0	55.0
`scan_object`	5.0	0.0	10.0	0.0	10.0	10.0
`shake_bottle`	100.0	100.0	100.0	90.0	100.0	100.0
`shake_bottle_horizontally`	100.0	100.0	100.0	95.0	100.0	100.0
`stack_blocks_three`	80.0	90.0	100.0	80.0	85.0	45.0
`stack_blocks_two`	95.0	100.0	100.0	85.0	95.0	85.0
`stack_bowls_three`	80.0	75.0	90.0	65.0	65.0	70.0
`stack_bowls_two`	100.0	95.0	95.0	85.0	90.0	70.0
`stamp_seal`	65.0	60.0	70.0	45.0	65.0	15.0
`turn_switch`	25.0	40.0	35.0	15.0	35.0	45.0
Mean	73.4	71.3	73.5	57.8	69.9	48.7

Limitations

The model is trained on clean scenes only, so Hard and Clean2Rand measure out-of-distribution robustness.
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}
}