Collaborative Research: CISE: Large: Executing Natural Instructions in Realistic Uncertain Worlds

For robots to fluidly operate as human assistants they must be able to take natural language instructions from humans and act to achieve those instructions in complex, uncertain environments. While there are commodity robots that are physically capable of carrying out a wide variety of useful instructions, current artificial intelligence (AI) frameworks are not able to fully understand and autonomously execute most of those instructions.

Rather, current AI techniques for robot instruction following have generally been limited to highly constrained, unrealistic environments and instruction formats that are unnatural, brittle, and rigid. The overarching goal of the project is to study the fundamental AI principles that enable robots to reliably execute natural instructions in realistic, uncertain worlds.

The project has the potential to dramatically increase the physical labor available to society, without increasing the amount of human labor, by enabling typical human workers to direct semi-automated robots for a multitude of mundane tasks. This will only be possible if the machines are easily instructable in natural environments by humans who only require minimal specialized training.

This envisioned labor multiplier is extremely relevant given the need for increased capacity to build physical infrastructure in the US. This includes, for example, new and upgraded public infrastructure such as bridges and energy systems, as well as efficient construction of affordable housing. These same advances will also result in broader impacts to other parts of the economy, such as logistics, healthcare, household assistants.

The project will also contribute to education and outreach through K-12 initiatives, undergraduate research experiences, and recruiting of underrepresented graduate student talent.

The project will design and develop a novel integrated framework for embodied AI agents that is comprised of synergistic advances in computer vision, language understanding, world modeling, planning, and control. The framework will evolve over a staged plan of increasing capabilities, starting with step-by-step instruction execution and progressing to executing general types of goal-oriented instructions.

The research will test and demonstrate the framework in both physically-realistic simulation environments and real-world environments using commodity robots. In addition, user studies will be conducted at each capability stage to focus the work toward end-user utility. Central to the framework is a new knowledge structure for spatio-temporal scenes, the multi-modal entity map (MEM), which is updated based on vision and language and used for both planning and skill execution.

The research will study new ideas in 3D vision and language understanding for continually maintaining the MEM based on realistic inputs in a way that captures uncertainty in the environment. The project will also study a new approach to low-level full-body control for robot skills, inspired by recent successes in language modeling, that facilitates both modular skill learning and knowledge sharing.

Finally, the project will advance automated planning capabilities by studying new ideas for learning dynamics models over the MEMs, which are used by a novel approach to high-level skill planning based on dynamics-conditioned language models. Importantly all of these innovations will be developed in a synchronized way to allow for rigorous testing and demonstration of the integrated framework.

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.