AI News Feed
Market watch
Engadget: Task Manager Helps Spot PC Background Processes That Drain ResourcesLocal Hubs and Backup Power Keep Smart Homes Running When Internet FailsEngadget Lists Ways to Repurpose, Sell or Recycle Old PlayStation ConsolesFederal Judge Blocks Utah's Anti-VPN Age-Verification LawDeepX pursues first-mover strategy in physical AI with Intel, AMD partnershipsNorwegian Army Rental Cars Stored Dashcam Footage of Soldiers, Report Says
Engadget: Task Manager Helps Spot PC Background Processes That Drain ResourcesLocal Hubs and Backup Power Keep Smart Homes Running When Internet FailsEngadget Lists Ways to Repurpose, Sell or Recycle Old PlayStation ConsolesFederal Judge Blocks Utah's Anti-VPN Age-Verification LawDeepX pursues first-mover strategy in physical AI with Intel, AMD partnershipsNorwegian Army Rental Cars Stored Dashcam Footage of Soldiers, Report Says
Robotics

Kyutai Releases REKI, a 2.4B Vision-Language-Action Model With Cached Memory for Robot Manipulation

Kyutai releases REKI, a 2.4B vision-language-action robot model trained on public data only, with cached memory and open weights.

AI News Feed · 2026-10-18 07:42

REKI pairs a Qwen3.5-2B vision-language backbone of 2.2 billion parameters with a 200-million-parameter flow-matching action expert. The expert generates action chunks in 10 Euler integration steps; robot state and flow time enter through adaptive RMSNorm, and each robot-state layout gets its own small state MLP. The model ships in BF16 under CC BY-NC-SA 4.0, with no quantized variants announced.

Most VLAs act on the current frame alone. REKI places memory inside the action expert instead: 32 learned register tokens precede each frame’s action tokens, cross-attend the encoded prompt and views of that frame, and self-attend registers from earlier frames. The registers never read action tokens, so they depend only on observations, not flow time or robot state. They are computed once per frame and cached in a rolling window covering the last three frames, one second apart. Kyutai says reading cached registers adds 0.1 ms to a 21.8 ms action chunk, the cache uses about 2 MiB of GPU memory per past frame, and training packs a three-frame window into one batched pass.

Kyutai says the memory fixes a failure mode in which robot arms oscillate between two poses. On RoboTwin, oscillating episodes fall from 18.6% to 4.1% with memory, a change Kyutai attributes to perceptual aliasing.

REKI was pretrained once and then finetuned separately per benchmark. It scores 90.4% on LIBERO-Plus, 65.8% on RoboTwin Clean2Rand and 38.4% on RoboCasa365. It also reports 97.6% across four LIBERO suites and 85.9% and 84.6% on RoboTwin 2.0 Easy and Hard. Kyutai says REKI is best among VLAs pretrained on public data on LIBERO-Plus and RoboTwin Clean2Rand. Its RoboCasa365 score trails Xiaomi-Robotics-1 at 57.4%, though it is ahead of RLDX-1 at 36.0% and Qwen-RobotManip at 35.9%.

In Kyutai’s comparison, Qwen-RobotManip, listed at 4.7 billion parameters, scores 89.0% on LIBERO-Plus, 68.5% on Clean2Rand and 35.9% on RoboCasa365. π0.5, listed at 3.3 billion parameters, scores 85.7%, 63.8% and 16.9%, while ABot-M0 at 4.6 billion parameters scores 80.5% and 55.7% on the first two benchmarks. Kyutai notes that the Qwen-RobotManip report lists π0.5 at 84.4% on LIBERO-Plus. Kyutai also says only models about twice REKI’s size beat it on three out-of-distribution benchmarks, while its RoboCasa365 result remains below Xiaomi-Robotics-1.

REKI’s training mixture contains 1.0 billion frames and 12.2 million vision-language samples. Real-robot teleoperation supplies 639.8 million frames, simulation 193.9 million and egocentric human video 170.9 million; sources include DROID, BridgeData V2, AgiBot World, EgoDex and EgoVerse. Pretraining ran for 200,000 steps on 32 H100 GPUs. Kyutai kept each dataset’s native control convention instead of remapping actions. Kyutai says pretraining helps most out of distribution, adding 13.8 points on Clean2Rand and 30.3 points under clutter; memory adds 1.5 points on LIBERO-Plus and 5.4 on RoboCasa365.

Six checkpoints are listed: the pretrained base and finetunes for LIBERO, RoboTwin in two variants, RoboCasa365 and RoboCasa-GR1. Real-robot deployment uses the LeRobot async-inference gRPC protocol. On a UR5 arm, REKI was finetuned for 3,000 steps, about one hour, on 50 demonstrations per task. Benchmarks were run on one NVIDIA H100, with 21.8 ms per action chunk when compiled, from two 256×256 cameras.