A new approach to training robots may soon enable them to perform complex tasks on job sites with greater efficiency, according to a study from the Massachusetts Institute of Technology (MIT). Researchers at MIT’s Improbable AI Lab, part of its Computer Science and Artificial Intelligence Laboratory (CSAIL), have developed a training method that combines three types of instructions—language-based, visual, and action-based—to teach robots how to complete multi-step tasks with higher success rates than traditional methods.
This training framework, named Compositional Foundation Models for Hierarchical Planning, or HiP for short, has the potential to impact not only construction but also manufacturing and even household chores. By using the three types of instruction, HiP enables robots to achieve their “long-horizon goals,” which are large objectives that require completing several smaller steps. These are tasks that humans find straightforward, like making a sandwich, brewing tea, or organizing materials by appearance.
HiP’s system operates on a hierarchical basis, allowing each component to be pre-trained on various datasets, including non-robotic information. This modular approach means different models leverage different types of internet data, which together improve robotic decision-making. The process incorporates “iterative refinement,” where each step builds upon the previous one, taking in feedback to create a more practical and effective plan.
The potential for robots in construction is considerable. Currently, robots are often limited by hardware constraints, yet they excel at heavy lifting and moving large objects. While humans are better at tasks requiring fine motor skills, robots can handle the physical labor that involves heavy lifting or moving objects to places inaccessible to cranes.
This blending of AI and robotics is a growing trend in construction, and while robots are not yet commonplace on job sites, the technology is rapidly advancing. HiP is just one example of AI-driven robotic training. Another instance is Toyota Research Institute’s use of a generative AI technique called “diffusion policy” to train their robots for complex tasks.
The impact of AI in construction extends beyond robots. AI can also streamline operations by automating tasks like email management, searching for information about change orders, or even automating design processes. As the technology develops, the construction industry is poised to benefit significantly from these advancements.
