This study introduces TS-Mask VLA, a vision-language-action (VLA) framework for robotic manipulation. It aims to comprehend natural language instructions and visual observations, generating and executing corresponding actions. Current autoregressive token-based action generation often simplifies action generation to next-token prediction, lacking explicit modeling of spatiotemporal structures, which limits performance in long-horizon and complex scenarios.
The TS-Mask VLA framework is built upon two key components:
-
Discrete Diffusion Action Expert: Equipped with a Bridge Attention mechanism, enabling multi-layer conditioning from the VLM, which enhances accuracy and stability in action generation.
-
Temporal-Spatial 2D Masking Strategy: This strategy for discrete action tokens strengthens the model's understanding of cross-time dependencies and inter-dimensional coupling, leading to structurally consistent action sequences.
Extensive experiments on simulation benchmarks and real-world tasks show that TS-Mask VLA achieves a 95.7% average success rate on LIBERO with only 0.5B parameters, significantly outperforming larger models. On CALVIN, it achieves the best average sequence length of 4.19 and demonstrates strong long-horizon performance. Comprehensive analyses and ablation studies further validate the effectiveness of our design.
Blogger's Review: The TS-Mask VLA model significantly enhances robotic performance in complex tasks through its innovative temporal-spatial masking strategy and efficient multi-layer conditioning mechanism. It showcases the tremendous potential of VLA models in practical applications, warranting further attention and research.