CAREER: Abstraction Barriers for Embodied Algorithms

Suny At Buffalo, Amherst NY

Investigators

Abstract

Computers have fundamentally changed society by being able to solve ever more complex problems quickly and reliably. One key to their utility is that programmers can quickly develop algorithms to solve new problems as by using well-understood mathematical abstractions of the underlying computational machinery. Ideally, modern robots should behave in the same way. They should perform useful tasks for people, and they should be easy to program so that roboticists can quickly adapt them to solve new problems as they arise. However, in practice, the process of developing robots is much more convoluted when they operate in unstructured, i.e., real-world, environments and requires iterative adjustments of the robot and its programming. The problem is that physical interaction is difficult to model, so that a robot action can inadvertently change the state of the world, sometimes directly causing accidents or causing problems in future robot-world interactions. This project addresses this problem in the context of robot construction by developing representations of the world state that robots can reason about and use for planning. These allow programmers to treat robots and embodied algorithms and to make robots that reliably operate when modifying the environment and building structures. This algorithmic view of robot-environment interaction is one approach to accomplishing the urgent societal need for innovative construction techniques that are adaptable and scalable and will help humans adapt to changing societal and infrastructure needs. This project develops abstraction barriers with different amounts of physical detail, which can be combined to produce robots that reliably work in highly unstructured environments. Project outcomes will impact several fields including 1) assemblies with pre-fabricated materials, 2) mixed-material assemblies with found objects, 3) embodied algorithms, 4) plan synthesis, and 5) probabilistic guarantees on construction processes. Specifically, this involves partial orders to encode motion, reachability, and mechanical assembly constraints with a variety of rigid materials; and geometric abstractions that can be used to provide correctness guarantees for robots that interact and modify highly irregular environments by using composable geometric primitives. This work can be directly useful for commercial construction robots. The proposed research comprises complementary ways in which physical constraints are represented, and will enable transformative cyber-physical systems able to reliably modify unstructured environments This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.