The Governing Council of the University of Toronto

Purpose

Automating the search through the chemical space of an estimated 10^20 to 10^60 drug-like molecules for a subset that have desired characteristics and features will be a major enabler for materials and drug discovery, chemical synthesis and medicine-by-design. However, research into fully autonomous chemistry labs that close the loop of i) optimal experiment design, ii) experiment execution (synthesis), and iii) observation and evaluation (characterization), are still in their infancy. A major challenge is that while today's robots have made significant strides in manipulating rigid objects, they lack the perceptual and dexterous capabilities needed to work directly with fluids and powders, which are omnipresent in chemistry labs. This is a critical research area as, in general, perceiving and manipulating fluids, powders and non-rigid materials is recognized as a longstanding challenge in the robotics community. Within the context of automated chemistry in human-centric labs, the main technical difficulties are the perception transparent and non-rigid materials, along with the lack of fast and accurate physical simulators for liquids, powders, and granular media. Existing physical simulators, for example, are either real-time or accurate, but not both, which makes planning and controlling the robots via visual feedback challenging and error prone. In this research, the project aims to develop machine learning methods to address the perceptions, modeling and simulation challenges associated with transparency, powders, liquids, and granular materials. The project seeks to then utilize these methods to develop learning-based control and perception methods that will enable general-purpose robot arms to manipulate non-rigid martials that commonly appear in chemistry labs. The experimental results will demonstrate fundamental chemistry operations such as pouring with different glassware and material types. The experiments will be carried out within Acceleration Consortium, University of Toronto, and NRC labs, showing reproducibility outside of a single lab. The research will complement the NRC PI’s existing AI4D projects on robotic chemists, which are focused on the transfer of lab items, but not pouring and material dynamics