TBD-VLA: Temporal Block Diffusion Vision Language Action Model

2026-06-05

AI Summary

This work addresses the high inference latency and poor temporal structure modeling in existing discrete vision-language-action (VLA) models. The authors propose a temporal block diffusion mechanism that partitions action sequences into temporal blocks, enabling parallel decoding within each block via masked discrete diffusion while preserving autoregressive generation across blocks to maintain temporal consistency and improve efficiency. The approach further supports asynchronous execution of action blocks through temporal inpainting, facilitating real-time chunked generation. By unifying autoregressive and parallel decoding paradigms, the method achieves significantly superior performance over current approaches in both simulated and real-world manipulation tasks, yielding faster and more temporally coherent discrete VLA models.

Abstract

Discrete Vision-Language-Action (VLA) models typically formulate action generation as next-token prediction over discretized action spaces, conditioning each token autoregressively on prior context. While effective, this paradigm incurs high inference latency and largely ignores the temporal structure inherent in action trajectories. Recent efforts introduce parallel decoding to improve efficiency, enabling faster inference, but lack explicit mechanisms for modeling token dependencies. We introduce TBD-VLA, a discrete token-based VLA framework that incorporates block diffusion to enable temporal action generation. We partition action sequences into temporal blocks and perform masked discrete diffusion within each block, while maintaining autoregressive generation across blocks. This design unifies temporal autoregression and parallel action decoding, achieving both strong temporal coherence and improved inference speed. In addition, the explicit temporal modeling enables asynchronous execution of action chunks (e.g., Real-Time Chunking) via temporal in-painting. TBD-VLA significantly outperforms prior VLA approaches in both simulation and real-world manipulation tasks, offering a scalable path toward fast, temporally aware, discrete VLA models. Project webpage: https://tbd-vla.github.io/

Problem

• Vision-Language-Action
• Temporal Structure
• Discrete Action Generation
• Inference Latency
• Token Dependencies

Innovation

• Temporal Block Diffusion
• Discrete Vision-Language-Action
• Parallel Action Decoding
• Temporal Autoregression
• Real-Time Chunking