Reproduce “TBD-VLA: Temporal Block Diffusion Vision Language Action Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sung-Wook Lee, Xuhui Kang, Yen-Ling Kuo
TBD-VLA is a Vision-Language-Action model that integrates temporal block discrete diffusion to efficiently generate robot actions, achieving both fast inference and strong generalization. The model demonstrated a 97.7% average success rate on LIBERO test suites and a 15.1% improvement over the next best method on LIBERO-Plus benchmarks, alongside competitive real-world performance and reduced inference latency to 0.086s.
Key insights
- Unifying parallel denoising of tokens within temporal blocks with autoregressive generation across these blocks effectively balances inference speed and temporal coherence for discrete action sequences.
- A VLM backbone, when trained with a temporal-level token shift and custom attention mask for block discrete diffusion, can implicitly model complex temporal dependencies within action trajectories.
- Asynchronous inference mechanisms like Real-Time Chunking (RTC), supported by the model's temporal in-painting capability, can significantly mitigate latency issues in closed-loop control by planning future actions while executing current ones.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.