Beyond Traditional Planning: A Discrete-Continuous Framework for Robotic Task Automation

The high-level research goal is to develop a real-time, adaptable robotics framework to support healthcare staff in eldercare facilities and hospitals. This framework will utilize both high-level discrete reasoning and low-level continuous motion reasoning to perform complex, multi-step tasks, such as meal preparation or diagnostics testing.

The central challenge is integrating discrete and continuous variables in a way that is computationally feasible. The proposed approach involves two modules: a high-level module for discrete reasoning and a low-level module for continuous motion reasoning. The high-level module will use inputs such as the robot’s goal, current state, environment description, and legal actions, to generate a set of actions to satisfy the goal. This set of actions will then be passed to a Large Language Model (LLM) which will select the most appropriate action based on the context. The selected action will then be passed to the low-level continuous module, which will use non-linear optimization algorithms to compute the necessary motions to complete the action.

The proposed framework will be evaluated against current task and motion planning approaches, with metrics including planning time, task completion time, adaptability, error rate, and optimality of actions. In addition, the system will be deployed in a real-world medical environment in collaboration with Yale Medical School.

The broader impacts of this research are substantial, with potential applications including assistance for healthcare workers in under-resourced clinics, expedited search and rescue missions in disaster areas, and replacing humans in hazardous tasks such as waste management, mining, or nuclear reactor maintenance. This research aims to improve working conditions and quality of life for many individuals, while also making cutting-edge technology accessible and beneficial to the broader public.