Reproduce “Visual Planning: Let's Think Only with Images” on OpenResearch

About this paper

Yi Xu, Chengzu Li, Han Zhou, Xingchen Wan, Caiqi Zhang, Anna Korhonen, Ivan Vulić

Researchers from the University of Cambridge and Google introduced Visual Planning, a paradigm enabling artificial intelligence models to perform multi-step planning directly through sequences of images, bypassing text-based reasoning. Their VPRL framework, utilizing a two-stage reinforcement learning approach with a large vision model, achieved superior performance on visual navigation tasks, outperforming language-based methods by a notable margin (e.g., 91.6% Exact Match on FROZENLAKE) and demonstrating enhanced robustness to complexity and out-of-distribution scenarios.

Key insights

• Models can perform effective multi-step planning purely through visual representations, providing a viable alternative to language-based reasoning for "vision-first" tasks.
• Explicitly avoiding textual mediation for visual reasoning helps circumvent the 'modality gap,' leading to more accurate and robust planning in visually complex environments.
• A two-stage reinforcement learning approach, combining supervised policy initialization for exploration with GRPO for outcome-based optimization, is effective for training generative vision models in sequential planning.

Summary by alphaXiv.