SBIR Phase I: Developing an Automated Outbound Packing System

Gridiron Robotics Llc, Chalfont PA

Investigators

Abstract

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project is to enable the fast and efficient loading of parcels into shipping containers ranging from small delivery vans to maritime shipping containers. This project will focus on demonstrating the feasibility of an algorithmic approach and robotic development. The technology is a step towards creating a fully autonomous system with expanded robotics capability to further enhance the efficiency and speed of outbound shipping for customers. Over 3,000 parcels are shipped every second. However, in the U.S., one out of every four trucks is empty, two are less than 50% filled, and only one is filled over 50% capacity. Initial projections indicate that the technology under development could decrease trucking costs by 20%, reduce loading costs by 70-80%, and decrease loading time by 30%, all while meeting the demands of peak shipping seasons. Overall, increasing the density of parcel shipping will reduce greenhouse gas emissions (400 tons/per truck/per year), reduce traffic congestion, and enable smaller businesses to compete with large organizations by reducing their logistics and shipping operating costs. This Small Business Innovation Research (SBIR) Phase I project will focus on advancing a bin packing algorithm to minimize void space in outbound shipping containers. The 3-Dimentional Bin Packing Problem (3D-BPP) is a classic Nonlinear Programming (NP)-hard problem that has been studied for decades. To solve the problem, an effective and easy-to-implement constrained, quantum accelerated, deep reinforcement learning model is being developed. Monte Carlo Tree Search is an unsupervised, heuristic search algorithm technique in which the learning agent learns to predict the expected value of a variable occurring at the end of a sequence of states. Deep reinforcement learning (DRL) extends this technique by allowing the learned state-values to guide actions which subsequently change the environment state. A proof-of-concept assessment showed that the learned strategy meaningfully outperforms the state-of-the-art methods. Outcome success metrics for this project are >90% utilization rate, sub 24 hours of model training time, and >2500 parcels/hour for any given data set. This foundation will be expanded by integrating many unique box sizes, exploring model performance in the face of broader circumstances (e.g., lookahead and stacking parameters, General Processing Unit (GPU) vs quantum training), and developing of a robotic gripper to enact algorithmic output. 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.