Accelerating Laboratory Automation Through Learning Tool Morphology For Robotic Chemists

The COVID-19 pandemic and current climate crisis have created an urgency for our societies to find a radically different way towards discovering new materials for more resilient, flexible pharmaceutical manufacturing and achieving net zero faster. Robotic chemists are increasingly being used in laboratories to accelerate materials discovery and alleviate scientists from tedious manual tasks. The development of robot chemists has been mainly driven by a blend of human intuition and classical engineering design. This has led to remarkable success for a variety of experiments, but the general design principles and formats of adapting robotic systems have now somewhat stagnated. For our application domain, the ideal robot does not yet exist; these generalised robotic systems are not up to the challenge of being used across heterogeneous experiments which currently are only achievable with human dexterity and skill. The dominant paradigm has been to develop multi-functional robotic platforms with predominantly industrial parallel grippers that can be deployed across a variety of tasks and environments. The main bottleneck is, and has always been, the robot’s end-effector which requires iterative manual design, prototyping and adaptation to each experimental workflow; this makes overall deployment in-flexible, time consuming and costly. It takes months, and years, of manual efforts to adapt state-of-the-art robotic chemists to new tasks, materials and instruments. Humans, and in particular human scientists, carry out everyday tasks using tools that have been designed for highly dexterous hands. The sheer diversity of tools used daily underscores the fact that the human hand alone is often insufficient for the complexity of most tasks. Tools embody the solution to addressing manipulation challenges by extending reach, adjusting force direction and magnitude, and overcoming physical limitations. While we have successfully automated some chemistry lab tasks using human tools or custom-designed 3D-printed tools based on task requirements, our works have highlighted a mismatch between the design of tools and the limited dexterity of robotic chemists’ end-effectors. We propose to change this by reconceptualising how tools are optimised for chemistry lab automation tasks such that they are effective with the robot’s embodiment and behaviour. Our research vision is to achieve this through data-driven co-design of tools for acquiring robotic manipulation skills capable of handling novel materials (e.g., fluffy powders, hygroscopic crystals, tacky materials), potentially with unknown properties, across different chemistry experiments. We will then autonomously design, produce and test a suite of optimised robotic tools for the designated laboratory tasks using an end-to-end production and evaluation unit. In collaboration with our industrial stakeholders, we will experimentally validate our methods across material-handling manipulation tasks e.g., sample scraping and powder weighing across at least 2 branches of chemistry with a balance of task-performance and safe, dexterous embodiment (pharma and materials chemistry; e.g., for batteries and clean technologies). The end result will be the construction of ‘robot chemists’ that are closer towards achieving autonomous chemical synthesis; a highly dexterous task that requires multiple tools. This project will undoubtedly challenge current methods of deploying robot scientists in labs; by demonstrating the feasibility of adapting human-designed tools for collaborative use, we will pave the way for a new era of optimised autonomous chemistry labs that will directly benefit the public because fusing robotics, chemistry and AI is the only competitive way to tackle the largest global challenges of our societies, from clean energy to public health.