An Adversarial Inverse Reinforcement Learning Approach to an Industrial Robotic Pick-and-Place

Abstract

Industrial automation cells are often engineered to be highly repeatable, yet many pallet-handling tasks remain over-constrained because key sources of variability are difficult to model explicitly. In production environments, pallets may remain in service for years and become worn, chipped, warped, or cracked. These changes introduce unpredictable contact behavior during grasping and placement through variations in friction, compliance, and seating. While machine vision is commonly used to estimate pallet pose and compensate for offsets, vision-based solutions can be sensitive to lighting and surface appearance and can increase cycle time due to image capture and processing.This thesis investigates an alternative approach that reduces reliance on non-value-added sensing by learning robust pick-and-place behavior from demonstrations. A physics-based palletizing work cell is developed in simulation, and expert trajectories are generated using a traditional state-machine strategy. Adversarial Inverse Reinforcement Learning (AIRL) is then used to learn a reward function from these demonstrations using a discriminator/reward network coupled with a policy generator. The AIRL stage demonstrates the ability to learn a reward function from demonstrated trajectories. A second stage of learning is applied, in which AIRL’s learned reward and initial policy are refined and added too with forward reinforcement learning using additional rewards to improve final state action and robustness under pallet worn pallets.