Fast-in-Slow: A Dual-System Foundation Model Unifying Fast Manipulation within Slow Reasoning

TL;DR

FiS-VLA presents a unified dual-system VLA model that embeds a fast execution module (System 1) inside a pretrained VLM-based slow reasoning system (System 2). Through a dual-aware co-training objective combining diffusion-based action generation with autoregressive reasoning, FiS-VLA achieves state-of-the-art manipulation performance while maintaining high control frequencies (up to 117.7 Hz with action chunking) and strong generalization to unseen objects, backgrounds, and lighting. The approach leverages heterogeneous modalities and asynchronous operation to balance fast reflexive control with deep multimodal reasoning, enabling robust single- and dual-arm manipulation in simulation and real-world settings. These results have practical implications for responsive, generalizable robotic manipulation in dynamic environments.

Abstract

Generalized policy and execution efficiency constitute the two critical challenges in robotic manipulation. While recent foundation policies benefit from the common-sense reasoning capabilities of internet-scale pretrained vision-language models (VLMs), they often suffer from low execution frequency. To mitigate this dilemma, dual-system approaches, inspired by Kahneman's theory, have been proposed to leverage a VLM-based System 2 model handling high-level reasoning and a separate System 1 action model ensuring real-time control. However, existing designs maintain both systems as separate models, limiting System 1 from fully leveraging the rich pretrained knowledge from the VLM-based System 2. In this work, we propose Fast-in-Slow (FiS), a unified dual-system vision-language-action (VLA) model that embeds the System 1 execution module within the VLM-based System 2 by partially sharing parameters. This innovative paradigm not only enables high-frequency execution in System 1 but also facilitates coordination between the reasoning and execution components within a single foundation model of System 2. Given their fundamentally distinct roles within FiS-VLA, we design the two systems to incorporate heterogeneous modality inputs alongside asynchronous operating frequencies, enabling both fast and precise manipulation. To enable coordination between the two systems, a dual-aware co-training strategy is proposed that equips System 1 with action generation capabilities while preserving System 2's contextual reasoning representation. For evaluation, FiS-VLA outperforms previous state-of-the-art methods by 8% in simulation and 11% in real-world tasks in terms of average success rate, while achieving a 117.7 Hz control frequency with action chunk set to eight. Project web page: fast-in-slow.github.io.

Figures (10)

• Figure 1: Overview of FiS-VLA. (a) Unlike previous dual-system VLA methods zhang2024hirtbu2025synergisticgeneralizedefficientdualsystem that attach a separate policy head as System 1, FiS-VLA (b) repurposes the final transformer blocks of an intact VLM as System 1, while retaining the full model for System 2 reasoning. Under this paradigm, FiS-VLA achieves superior performance and high-frequency control, as shown in (c) and (d).
• Figure 2: Framework of FiS-VLA. FiS-VLA leverages an intact VLM for System 2 reasoning while repurposing the final transformer blocks of the LLM for System 1 execution module. System 2 handles low-frequency inputs such as 2D images and language instructions and produces intermediate latent features that serve as conditioning information for System 1. Instead of being conditioned solely on these periodically updated high-level representations, System 1 processes high-frequency inputs including 3D point clouds, 2D images, and robot states to produce stable and responsive actions. For joint optimization, we introduce a dual-aware co-training strategy that combines a diffusion denoising objective with an autoregressive objective which enables FiS-VLA to support fast action generation while retaining System 2’s multimodal reasoning capabilities.
• Figure 3: Ablation study. We investigate the impact of (1) the parameters of System 1’s shared blocks within System 2, (2) different modality inputs to System 1, and (3) the operating frequency ratio between the two systems on final manipulation success rates.
• Figure 4: Real-world assets and camera configurations. We present visualizations of the real-world assets and camera setups used in the Agilex and AlphaBot dual-arm robot tasks, respectively.
• Figure 5: Ablation studies on action chunk size and input variants of FiS-VLA. (Left) Impact of different action chunk sizes on success rate and inference speed. While increasing action chunk size leads to improved inference speed, success rate remains relatively stable. (Right) Comparison of success rates among FiS-VLA and its input variants, showing FiS-VLA achieves the best performance.