Home › Robotics › FAST: Efficient Action Tokenization for Vision-Language-Action Models

FAST: Efficient Action Tokenization for Vision-Language-Action Models - AI Robotics Tool

Overview

FAST (FAST+) is a universal action tokenizer that maps robot action sequences into dense, discrete tokens. It provides a pre-trained tokenizer and tooling to train custom tokenizers for autoregressive vision-language-action model training.

Key Features

  • Maps robot action sequences to dense, discrete tokens
  • Pre-trained FAST+ tokenizer available for immediate use
  • Tools and scripts to train custom tokenizers on user data
  • Designed for autoregressive vision-language-action model training
  • Produces compact token sequences suitable for sequence modeling

Ideal Use Cases

  • Tokenizing robot action datasets for model training
  • Training autoregressive vision-language-action models
  • Fine-tuning tokenizers on custom robot behavior data
  • Compressing action sequences for sequence modeling and evaluation

Getting Started

  • Visit the Hugging Face repository: https://huggingface.co/physical-intelligence/fast
  • Load the provided pre-trained FAST+ tokenizer into your pipeline
  • Prepare robot action sequences in the repository's expected format
  • Run the supplied training scripts to train a custom tokenizer
  • Convert tokenized sequences into inputs for autoregressive model training

Pricing

Not disclosed; repository available on Hugging Face.

Key Information

  • Category: Robotics
  • Type: AI Robotics Tool
Visit Official Website

Related Tools

LeRobot

A comprehensive suite of models, datasets, and tools for real-world robotics, providing end-to-end learning resources in PyTorch for robotics AI applications.
Genesis

Genesis is a physics simulation platform and generative data engine for general-purpose robotics and embodied AI learning. It features a universal physics engine, ultra-fast robotics simulation, photorealistic rendering, and a modular generative framework that converts natural language prompts into various data modalities.
NVIDIA Isaac GR00T N1

An open foundation model for generalized humanoid robot reasoning and manipulation. It accepts multimodal inputs (language and images) and uses a vision-language foundation model combined with a diffusion transformer head to denoise continuous actions, enabling robot control and fine-tuning for specific tasks and embodiments.
NVIDIA Isaac GR00T N1

The world’s first open foundation model for generalized humanoid reasoning and skills, aimed at advancing robotics systems and autonomous vehicle technology.
Reachy Mini

Reachy Mini is an expressive, open‐source robot designed for human-robot interaction, creative coding, and AI experimentation. Programmable in Python (with upcoming support for additional languages), it integrates with state-of-the-art Hugging Face models for speech, vision, and personality. Available as a kit in a lite or wireless version, it is intended for AI developers, researchers, educators, and hobbyists to build, test, and share real-world AI applications.
ai2thor

ai2thor is an open-source simulated environment platform for Visual AI. It provides high-quality, interactive indoor scenes with detailed household objects, multi-agent support, a wide range of actions, and diverse sensory data. The platform is designed to facilitate research in embodied AI, including common sense reasoning, navigation, and manipulation tasks.