| Property | Value |
|---|---|
| Author | physical-intelligence |
| Model Type | Action Tokenizer |
| Source | HuggingFace |
What is FAST?
FAST (Efficient Action Tokenization for Vision-Language-Action Models) is a revolutionary tokenizer designed specifically for robotics applications. It efficiently converts sequences of robot actions into discrete, dense tokens that can be used to train autoregressive Vision-Language-Action (VLA) models. The system includes FAST+, a universal action tokenizer trained on 1 million real robot action sequences.
Implementation Details
FAST is implemented as a HuggingFace AutoProcessor, making it easily accessible and integrable into existing workflows. The system operates on 1-second action "chunks" that are pre-normalized to a range of [-1...1] using quantile normalization. It supports batched inference for both encoding and decoding operations.
- Simple installation through pip (transformers and scipy packages)
- Supports batch processing of action sequences
- Automatic dimension handling during decoding
- Custom tokenizer training capabilities
Core Capabilities
- Efficient conversion of continuous action sequences to discrete tokens
- Universal tokenization across different robot setups
- Quick training of custom tokenizers on specific datasets
- Seamless integration with HuggingFace ecosystem
- Support for variable-length action sequences
Frequently Asked Questions
Q: What makes this model unique?
FAST+ is unique in its ability to handle a wide range of robot setups, action dimensions, and control frequencies through a universal tokenization approach. It's been trained on 1M real robot action sequences, making it robust and versatile.
Q: What are the recommended use cases?
The model is ideal for robotics applications requiring discrete representation of continuous action sequences, particularly in vision-language-action systems. It's especially useful for training autoregressive models and standardizing robot action data across different platforms.