Kyutai REKI VLA Model Uses Register Tokens to Cut Robot Oscillation
Kyutai’s new 2.4B model uses efficient memory registers to reduce robotic arm oscillation and improve manipulation scores on public benchmarks.
Key points
- Kyutai released REKI, a 2.4B parameter vision-language-action model for robot manipulation.
- The model uses 32 learned register tokens to cache visual memory, reducing oscillation.
- REKI achieved the top score on the LIBERO-Plus benchmark among public-data models.
Kyutai, alongside Inria and École des Ponts, launched REKI, a vision-language-action model designed for robot manipulation. According to MarkTechPost, this 2.4B parameter system aims to provide efficient memory capabilities for robotic tasks. All model weights are now available on Hugging Face for immediate access by developers and researchers.
What is the technical design of REKI?
REKI combines a Qwen3.5-2B vision-language model backbone with a 200M-parameter action expert. The action expert generates movement commands using flow matching integrated in 10 Euler steps. Robot state and flow time enter the system through adaptive RMSNorm, with each state layout assigned a small multi-layer perceptron for processing.
Unlike most vision-language-action models that process only the current camera frame, REKI incorporates a memory mechanism. This design allows the system to retain context from recent observations without the heavy computational cost associated with traditional memory methods.
How does the memory mechanism function?
The model places memory inside the action expert using 32 learned register tokens. These tokens cross-attend to encoded prompts and camera views for each frame. They also self-attend to registers from earlier frames, creating a continuous thread of visual context.
Registers do not read action tokens, meaning they depend solely on observations. This separation allows the system to compute registers once per frame and cache them in a rolling window, significantly improving efficiency.
What are the performance results?
On the LIBERO-Plus benchmark, REKI achieved a 90.4% success rate, outperforming Qwen-RobotManip at 89.0% and π0.5 at 85.7%. According to Kyutai, this is the best score among models pretrained on public data.
On RoboTwin Clean2Rand, the model scored 65.8%, trailing only Qwen-RobotManip at 68.5%. However, performance dropped to 38.4% on RoboCasa365, which remains behind Xiaomi-Robotics-1.
How does memory impact operational stability?
Kyutai reports that the memory feature addresses a specific failure mode where robotic arms oscillate between two poses. On RoboTwin benchmarks, episodes with oscillation fell from 18.6% to 4.1% when memory was enabled.
This improvement is attributed to reduced perceptual aliasing. The cached registers add only 0.1 ms to the action chunk generation time, which totals 21.8 ms when compiled on an NVIDIA H100.
Questions to ask your suppliers
Operations teams should evaluate whether new models introduce dependencies on specific hardware configurations or data sources. Security leads need to verify if training data includes sensitive internal information or if the model architecture exposes proprietary process details. Procurement should confirm that licensing terms allow for modification and internal deployment without external connectivity requirements. Engineering managers must assess the maintenance burden of integrating new AI components into existing robotic control loops.
What to do and how to stay safe: Kyutai
- Verify that model weights and inference code are validated against internal security policies before deployment.
- Monitor robotic control logs for unexpected oscillation or deviation from planned paths during testing.
- Ensure that camera feeds and action outputs are isolated from other network segments.
- Review data retention policies to confirm that cached visual frames do not persist longer than necessary.
Step-by-step guide: Browser Security Settings for Small Business Teams
General security guidance from the TechSec Feed newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What benchmarks did REKI perform best on?
REKI achieved the top score on LIBERO-Plus at 90.4% and was second on RoboTwin Clean2Rand at 65.8%.
How much GPU memory does the cache use?
The cache uses approximately 2 MiB of GPU memory for each past frame stored in the rolling window.
Is REKI pretrained on private data?
No, the model was pretrained only on public data before being finetuned for specific benchmarks.
