FAST: Efficient Action Tokenization for Vision-Language-Action Models (2501.09747v1)

Abstract: Autoregressive sequence models, such as Transformer-based vision-language action (VLA) policies, can be tremendously effective for capturing complex and generalizable robotic behaviors. However, such models require us to choose a tokenization of our continuous action signals, which determines how the discrete symbols predicted by the model map to continuous robot actions. We find that current approaches for robot action tokenization, based on simple per-dimension, per-timestep binning schemes, typically perform poorly when learning dexterous skills from high-frequency robot data. To address this challenge, we propose a new compression-based tokenization scheme for robot actions, based on the discrete cosine transform. Our tokenization approach, Frequency-space Action Sequence Tokenization (FAST), enables us to train autoregressive VLAs for highly dexterous and high-frequency tasks where standard discretization methods fail completely. Based on FAST, we release FAST+, a universal robot action tokenizer, trained on 1M real robot action trajectories. It can be used as a black-box tokenizer for a wide range of robot action sequences, with diverse action spaces and control frequencies. Finally, we show that, when combined with the pi0 VLA, our method can scale to training on 10k hours of robot data and match the performance of diffusion VLAs, while reducing training time by up to 5x.

• The paper introduces FAST, a DCT-based tokenization method that efficiently compresses high-frequency robotic actions.
• It enhances training efficiency by up to 5x compared to traditional techniques, improving policy convergence in VLAs.
• FAST achieves universal applicability across diverse robotic systems without retraining, streamlining model performance.

Analysis of "Efficient Action Tokenization for Vision-Language-Action Models"

This paper examines the challenge of action tokenization within the context of vision-language-action (VLA) models, with a focus on robotics applications. It introduces a novel method for tokenization, titled \ModelAcronym, which addresses the inefficiencies and inaccuracies associated with traditional tokenization techniques in high-frequency robotic action datasets.

Key Contributions

The primary contribution lies in the proposal of a new tokenization strategy based on the Discrete Cosine Transform (\ModelAcronym). This approach is motivated by the inadequacies of existing tokenization methods, particularly the na\"{i}ve binning scheme, which struggles with high-dimensional, high-frequency data such as those encountered in dexterous robotic manipulation tasks.

The paper introduces \ModelAcronym, a tokenization algorithm that employs the Discrete Cosine Transform for compressing robot actions into more informative discrete tokens. This novel compression-based strategy leverages byte pair encoding (BPE) to reduce redundancy, providing efficient and effective tokenization without the need for extensive model training seen in other methods like VQ-based tokenizations.

Moreover, the research demonstrates the applicability of this method across various robotic control systems with differing action spaces, showcasing a universal action tokenizer called \ModelUniversalAcronym. This allows for consistent performance across diverse robot morphologies and control frequencies without the need to retrain the tokenizer for each new setting.

Experimental Evaluation

The authors perform thorough empirical evaluations using two prevalent VLA architectures—$\pi_0$ and OpenVLA—across a range of real-world and simulated robotic scenarios. The study highlights significant improvements in training efficiency and policy effectiveness with \ModelAcronym tokenization compared to the na\"{i}ve binning approach. The compression performance also surpasses that of more complex learned tokenization approaches like FSQ, particularly for high-frequency tasks involving intricate manipulations.

Key results indicate that \ModelAcronym facilitates a reduction in training time by up to 5x while maintaining or surpassing existing models' performance on tasks. This reduction stems from enhancing the informativeness of action tokens, improving the convergence rate of the models during training.

Implications and Future Directions

The introduction of \ModelAcronym has significant implications for the scalability and efficiency of VLAs in robotic systems. It provides a straightforward, computationally efficient means of handling high-frequency action sequences, a previously challenging domain for autoregressive models. The universal applicability of \ModelUniversalAcronym further positions this approach as a pragmatic default for a wide range of robotic systems.

Looking forward, the methodology opens avenues for further refinement and exploration. Future work could delve into optimizing the inference speed of autoregressive models trained with \ModelAcronym tokenization, addressing the current limitation of slower inference times compared to diffusion-based approaches. Additionally, the research community might explore the integration of compression techniques with alternative non-autoregressive model architectures to potentially harness the benefits of both paradigms.

Overall, this paper contributes a practical and theoretically sound framework for enhancing the efficiency of VLAs in robotic applications, inviting further exploration into generalized and adaptive tokenization strategies in AI.