Robotics Engineering Practicum Presentation - Tejas Rane

Robots in a Pinch: Cart Pickup as a Case Study in Tight-Space Autonomy

Companies like Aethon Inc. develop mobile robotic systems to operate in busy, semi-structured indoor spaces, like hospitals, and perform logistical tasks such as delivering meals, transporting linens, and collecting trash. This thesis conducted in collaboration with Aethon, focuses on one of their more challenging operations - the autonomous pickup of utility carts - where the robot must reverse into a narrow gap between the cart's casters with sub-inch precision to initiate lifting. This task serves as a representative case for robot navigation in tightly constrained environments, demanding both accurate perception and agile control.

While a rule-based pipeline for cart pickup is already deployed in the robot, it relies on fixed and known properties of the carts.  The perception module depends on hand-tuned geometric heuristics, and the control system is open-loop. Eliminating the reliance on these fixed properties increases the variety of carts that can be handled as well as reduces the edge-case pickup failures.  To address these issues, this thesis investigates machine learning methods as a self-adapting, generalizable alternative.

Although this thesis focuses on cart pickup, the learning-based architecture is extensible to other high-precision tasks - including docking on charging stations, entering elevators, and maneuvering around obstacles. This work demonstrates how integrating machine learning can open new market opportunities with increased performance capabilities, improve system adaptability, reduce development overhead, and accelerate deployment in dynamic environments such as hospitals.

Advisor:  Professor Constantinos Chamzas (WPI)

Committee:  Professor Navid Dadkhah Tehrani (Adjunct teaching Professor, WPI) and
   George Lucas (Director of Software Engineering, ST Engineering Aethon Inc)

Zoom:  https://wpi.zoom.us/j/97498688024