MIT scientists develop AI system to improve robot planning
The method could help robots operate better in changing environments, with the system supporting autonomous driving and collaborative robotic assembly.
Researchers at MIT have developed a hybrid AI framework designed to improve how robots plan and perform complex visual tasks. The approach combines generative AI with classical planning software, allowing machines to analyse images, simulate actions, and generate reliable plans to reach a goal.
The system relies on two specialised vision-language models. One model analyses an image, describes the environment, and simulates possible actions, while a second model converts those simulations into a formal programming language used for planning.
Generated files are then processed by established planning software to produce a step-by-step strategy.
Testing showed a significant improvement compared with existing techniques. The framework achieved an average success rate of about 70 percent, while many baseline methods reached roughly 30 percent.
Performance remained strong in unfamiliar scenarios, demonstrating the system’s ability to adapt to changing conditions.
The method could support applications such as robot navigation, autonomous driving, and multi-robot assembly systems. Continued development aims to handle more complex environments and reduce errors caused by AI model hallucinations.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!