Reproduce “A Comprehensive Review of Generative Physical Artificial Intelligence” on OpenResearch

About this paper

Satyam Gaba, Krutiksinh Rana, Siva Sai, Vinay Chamola, Dusit Niyato

Generative Physical Artificial Intelligence (GPAI) is reviewed as a closed-loop approach that connects multimodal perception, reasoning, action generation, feedback, simulation, and physical control. The survey organizes the field into five complementary model classes—robot foundation models, vision-language-action models, large behavior models, diffusion policies, and world foundation models—while comparing their applications, reported benchmark results, and unresolved safety and deployment challenges.

Key insights

• The five model classes are complementary: foundation and vision-language-action models provide broad task understanding, diffusion policies generate smooth multimodal trajectories, large behavior models encode motion priors, and world models support prediction, planning, simulation, and rare-event generation.
• Reported results indicate improving generalization and control, including RT-1's 67% success across more than 700 manipulation tasks, RT-2's 62% success on unseen tasks versus 32% for RT-1 in the cited comparison, and an average 47% improvement attributed to Diffusion Policy across 12 tasks and four suites.
• The field is moving toward modular or hierarchical combinations rather than a single universal model, but benchmark performance remains an incomplete proxy for robustness under noise, embodiment changes, contact uncertainty, and safety-critical disturbances.

Summary by alphaXiv.