Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.22178

REKI: Registering Knowledge in VLAs

Juliette Marrie·Paul Pacaud·Adrien Ramanana Rahary·Shizhe Chen·Moritz Böhle
View original source ↗Read source document ↗Export metadata ↓

Abstract

We introduce REKI, a 2.4B vision-language-action model (VLA) pretrained exclusively on public data that sets a new Pareto frontier between performance and model size: on LIBERO-Plus, RoboTwin Clean2Rand and RoboCasa365, it is outperformed only by models about twice its size. REKI consists of a vision-language model (VLM) followed by a flow-matching action expert, and is pretrained on a curated mixture of public robot, egocentric human and vision-language data. Additionally, we endow it with a memory that gives the policy access to its recent past at negligible cost: a few register tokens in the action expert summarise the output of the VLM at each frame and persist in the expert's cache. This memory improves success in simulation and on real robots, and makes oscillations between poses about four times rarer. REKI uses a simple model design and training paradigm, and is optimised to be efficient at training and inference. We release the full code across data curation, training, inference and evaluation, deployment, and all pretrained and finetuned model weights: https://github.com/kyutai-labs/reki.

Explore related subjects

cs.ROcs.CVcs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Juliette Marrie, Paul Pacaud, Adrien Ramanana Rahary, Shizhe Chen, Moritz Böhle. 2026-10-15. REKI: Registering Knowledge in VLAs. https://arxiv.org/abs/2610.22178

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Path Planning on Multi-level Point Cloud with a Weighted Traversability Graph

This article proposes a new path planning method for addressing multi-level terrain situations. The proposed method includes innovations in three aspects: 1) the pre-processing of point cloud maps with a multi-level skip-list structure and data-slimming algorithm for well-organized and simplified map formalization and management, 2) the direct acquisition of local traversability indexes through vehicle and point cloud interaction analysis, which saves work in surface fitting, and 3) the assignment of traversability indexes on a multi-level connectivity graph to generate a weighted traversability graph for generally search-based path planning. The A* algorithm is modified to utilize the traversability graph to generate a short and safe path. The effectiveness and reliability of the proposed method are verified through indoor and outdoor experiments conducted in various environments, including multi-floor buildings, woodland, and rugged mountainous regions. The results demonstrate that the proposed method can properly address 3D path planning problems for ground vehicles in a wide range of situations.

cs.RO↗
arXiv2026

An Introduction to Zero-Order Optimization Techniques for Robotics

Zero-order optimization techniques are becoming increasingly popular in robotics due to their ability to handle non-differentiable functions and escape local minima. These advantages make them particularly useful for trajectory optimization and policy optimization. In this work, we propose a mathematical tutorial on random search. It offers a simple and unifying perspective for understanding a wide range of algorithms commonly used in robotics. Leveraging this viewpoint, we classify many trajectory optimization methods under a common framework and derive novel competitive RL algorithms.

cs.RO↗
arXiv2026

Drive As You Like: Multi-Head Diffusion with Reinforcement Learning for Personalized Driving

Personalized autonomous driving requires understanding the driving needs implicit in user requests. This is challenging when driving demonstrations lack style annotations and behavioral adaptation risks degrading planning quality. We propose M-Diffusion Planner with a dual-decoupling design: separating intent understanding from trajectory planning, and general planning capability learning from strategy alignment. A large language model infers driving needs and selects a strategy, which is executed by the corresponding head of a shared diffusion planner. Following imitation pretraining, strategy-specific rewards specialize individual heads without style annotations. We innovatively adapt Group Relative Policy Optimization using surrogate policy-ratio estimation and constrained head-specific updates to mitigate degradation of pretrained planning capabilities. Open-loop and closed-loop evaluations on nuPlan demonstrate distinct strategy-aligned behaviors and competitive planning performance. Further experiments assess intent-to-strategy mapping and collision risk, and analyze the effects of reward design and optimization on strategy alignment and planning quality.

cs.RO↗