What Is Fast-in-Slow (FiS-VLA)?

FiS-VLA is introduced in Fast-in-Slow: A Dual-System Foundation Model Unifying Fast Manipulation within Slow Reasoning (arXiv 2506.01953; submitted 2 June 2025, cs.RO) by Hao Chen, Jiaming Liu and 9 co-authors from The Chinese University of Hong Kong, Peking University, AI2Robotics and BAAI. The abstract names the model FiS-VLA, and the PDF is marked as a preprint under review. Read: the arXiv abstract page, the paper text through part of the ablation study, the official project page and the GitHub README.

FiS-VLA is third-party research. Firsthand is not part of it or affiliated with its authors, and makes no claim that any model was trained on Firsthand data. What the site states is custom egocentric collection of real human demonstrations, fast and on demand, to a written spec: 3D hand pose at 21 joints per hand, calibrated stereo depth and 6-DoF IMUs at 200 Hz, delivered in RLDS, LeRobot format or WebDataset with a buyer-owned commercial license. The only language data the site states is one task instruction per episode. Mapping any delivery to FiS-VLA’s inputs or action format is Not stated.

Per the abstract, foundation policies built on internet-scale pretrained VLMs gain common-sense reasoning but often run at a low execution frequency. Dual-system designs, inspired by Kahneman’s theory, pair a VLM-based System 2 for high-level reasoning with a separate System 1 action model for real-time control.

The authors argue that existing designs keep the two as separate models, which stops System 1 from fully using the pretrained knowledge in the VLM.

FiS-VLA keeps an intact VLM for System 2 and repurposes the final transformer blocks of its LLM as the System 1 execution module. The paper uses a 7B LLaMA2 backbone of 32 blocks, initialized from a Prismatic VLM, with SigLIP and DINOv2 vision encoders.

System 2 handles low-frequency 2D images and the language instruction and produces latent features that condition System 1. System 1 takes high-frequency inputs: 3D point clouds, 2D images and the robot state.

It is the paper’s joint objective. System 1 learns action generation with a diffusion denoising loss, while System 2 keeps its reasoning through an autoregressive next-token prediction loss, so adding fast control does not erase the slow reasoning.

The authors pretrain on open-source robotic datasets with more than 860K trajectories, then fine-tune on high-quality self-collected real-world and simulation data. They report a 1:4 frequency ratio between System 2 and System 1.

The abstract reports that FiS-VLA outperforms previous state-of-the-art methods by 8% in simulation and 11% in real-world tasks in average success rate, with a 117.7 Hz control frequency at action chunk eight. In the paper’s RLBench comparison of 10 tasks, FiS-VLA reaches a 0.69 mean success rate at 21.9 Hz with action chunk one on an NVIDIA 4090.

Real-world tests used two dual-arm robots, AgileX and AlphaBot. These are the authors’ results on their own setups; Firsthand does not reproduce them or rank models.

The site’s entries describe HybridVLA as combining diffusion and autoregressive action prediction in one LLM, RoboDual as a slow VLA generalist guiding a fast diffusion specialist, OpenHelix as an open-source dual-system model and short survey, and Hi Robot as a hierarchical VLA. The FiS-VLA README says its code is built on HybridVLA and Lift3D.

The paper contrasts FiS-VLA with dual-system methods that attach a separate policy head as System 1. How it relates to RoboDual, OpenHelix or Hi Robot specifically is Not stated in the sources read.

The paper trains on robot trajectories, with robot state and point clouds from depth maps, and mentions manually annotated sub-task plans during fine-tuning. On the site, the only language data is one task instruction per episode, and Firsthand captures human egocentric video rather than robot data.

The paper text read stops partway through the ablation study, and the appendices were not read, so remaining ablations, generalization results and hyperparameters are Not stated here. The abstract and project page give different control frequencies (117.7 Hz at action chunk eight, and 21.9 Hz without action chunking), so check which setting a figure refers to.

When a program needs real human egocentric demonstrations of tasks, environments or objects that public data does not cover, specified in writing. Firsthand does not build or train models.

Confirm the spec on a Spec Pilot, a small fast first batch. Where coverage is deep, the median from signed spec to first validated batch is 48 hours; where it is thin, three to four weeks. Billing is per validated hour.

What is Fast-in-Slow (FiS-VLA)?

A published dual-system vision-language-action model (arXiv 2506.01953) that embeds a fast System 1 execution module inside a slow VLM-based System 2 by partially sharing parameters. It is third-party research.

What does fast manipulation within slow reasoning mean?

Per the paper, the final transformer blocks of the LLM act as System 1 for quick action generation, while the full VLM stays as System 2 for slower reasoning. Both are one foundation model, not two.

How fast is FiS-VLA?

The authors report 117.7 Hz at action chunk eight in the abstract, and 21.9 Hz at action chunk one on an NVIDIA 4090 in their RLBench table. These are their results; Firsthand ranks no models.

Is FiS-VLA open source?

The official GitHub repository, CHEN-H01/Fast-in-Slow, holds the code, and GitHub lists an MIT License. Its README also says a pretrained checkpoint was released on 2025/07/08.

Does FiS-VLA use human video or language data?

In the sources read it trains on robot trajectories with 2D images, robot state and point clouds. Human egocentric video is not described. The only language data the site states is one task instruction per episode.

Does Firsthand data train FiS-VLA?

No claim is made. Firsthand is not affiliated with its authors, and the site does not state that any model was trained on Firsthand data.