MIT Develops Artificial Intelligence to Make Robots Packaging Experts

Some things are easy for humans but frustrating for machines. You might effortlessly stack groceries in a suitcase or arrange a table, but robots can be tripped up by the numerous constraints we handle without hesitation. A team at MIT has created a new generative AI tool that makes it easier to solve multi-step 3D manipulation tasks, ultimately replacing human packing and organizing work.

Robots can easily understand how to move objects from one place to another, especially if they are identical. The more diverse the objects, the more constraints there are. For example, when packing groceries, you don't want to put heavy items on top of light ones. If you're setting a table, you want to make sure the utensils are close together, but they should also be near the plates, and so on. Traditional AI for robots tackles these problems sequentially, designing partial solutions for each constraint and checking if they violate any other constraints. With only a few rules, most systems stagnate or produce suboptimal solutions.

MIT's new system is based on a generative AI approach in the form of a diffusion model. The tool, Diffusion-CCSP, consists of a model classifier trained for each individual constraint. This approach requires less training data compared to integrated models. These models work together to generate faster overall solutions, taking into account all necessary stacking and organizing rules.