CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models
TLDR
Introduces visual chain-of-thought reasoning into VLAs by autoregressively predicting future frames as goals, improving manipulation performance.
Reasoning
The paper presents a novel method for incorporating explicit visual reasoning into VLAs, achieving strong empirical results in both real-world and simulated tasks. However, it does not explicitly frame its contribution as a world model, and the abstract lacks detailed analysis of limitations or comparisons to world model approaches.
Read-first score
Read-first score 55.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 27.
Field roles
Rank sensitivity
Stability: volatile; rank range: 432.
Keyword Scores
Deep Analysis
Innovations
- Explicit visual chain-of-thought reasoning for vision-language-action models by autoregressively predicting future image frames as visual goals before generating action sequences.
- A unified 7B vision-language-action model (CoT-VLA) that jointly understands and generates both visual and action tokens.
Methodology
CoT-VLA incorporates explicit visual chain-of-thought reasoning by autoregressively predicting future image frames as visual goals, then generating a short action sequence to achieve those goals. The model is a 7B parameter VLA that processes and generates both visual and action tokens, leveraging pretrained vision-language models and diverse robot demonstrations for training. Evaluation is conducted on real-world manipulation tasks and simulation benchmarks, comparing against a state-of-the-art VLA baseline.
Key Results
CoT-VLA outperforms the state-of-the-art VLA model by 17% in real-world manipulation tasks and 6% in simulation benchmarks.