Robots are becoming more and more prevalent in our homes, offering assistance with various tasks. However, when faced with a new situation, these robots often struggle to figure out the most sensible actions to take. That’s where PIGINet comes in. Developed by researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), PIGINet is a system that uses machine learning to enhance the problem-solving capabilities of household robots.
Traditionally, robots go through an iterative process of task planning, where they attempt various task plans and refine their moves until they find a feasible solution. This can be time-consuming, especially when there are obstacles to consider. PIGINet eliminates task plans that can’t meet collision-free requirements, saving a significant amount of planning time. In fact, when trained on only 300-500 problems, PIGINet reduces planning time by 50-80 percent.
PIGINet is different from other systems because it doesn’t rely on predefined rules. Instead, it uses a neural network that takes in various inputs, such as plans, images of the environment, and the desired goal. The network then predicts the probability that a task plan can be refined to find feasible motion plans. By combining these different inputs, PIGINet can generate a prediction about the feasibility of a selected task plan.
To test PIGINet, the researchers created simulated kitchen environments with different layouts and tasks. They compared PIGINet against prior approaches by measuring the time taken to solve problems. PIGINet significantly reduced planning time, even in complex scenarios with longer plan sequences and less training data.
One of the challenges faced during the development of PIGINet was the lack of good training data. However, the team was able to overcome this by using pretrained vision language models and data augmentation techniques. This allowed PIGINet to show impressive plan time reduction, even on problems with unseen objects.
The goal of PIGINet is to create adaptable and practical household robots that can navigate complex and dynamic environments. The researchers also hope to refine PIGINet further to suggest alternate task plans and speed up the generation of feasible plans without the need for large datasets.
Overall, PIGINet offers a promising solution to the challenge of efficient decision-making in unstructured environments. It eliminates the need for oscillation between motion and task planning, making it more computationally efficient. With further refinement, PIGINet could revolutionize the way robots are trained and applied in various settings.
The research paper on PIGINet, authored by MIT’s Zhutian Yang and other collaborators, will be presented at the Robotics: Science and Systems conference in July. The project was supported by AI Singapore and grants from the National Science Foundation, the Air Force Office of Scientific Research, and the Army Research Office.