EVALUATING HUMANOID ROBOTS FOR INDUSTRIAL APPLICATIONS

Sammanfattning

Humanoid robots have gained increasing attention in industrial automation due to their potential to operatein environments designed for humans while performing repetitive and physically demanding tasks. Recent advancements in Vision Language Action (VLA) models have enabled robots to combine visual perception, language understanding and motion control within unified learning-based systems. This thesis investigates the implementation and evaluation of a humanoid robot for an industrial pick-and-place application usingstate-of-the-art VLA models.The work is based on the Unitree G1 humanoid robot equipped with Inspire FTP dexterous hands and teleoperated through an Extended Reality (XR) setup using a Meta Quest 3 headset. A complete pipeline for teleoperation, dataset collection, simulation, model training and real-world evaluation was developed. Demonstration datasets were collected both in simulation and on the physical robot using XR teleoperationand converted into the LeRobot framework format for training. Two VLA architectures, Gr00t and π0.5, were trained and evaluated on an industrial pick-and-place task provided by Volvo Construction Equipment.The results indicate that XR teleoperation can be used to generate datasets for humanoid imitation learning, and that a fine-tuned Gr00t policy can achieve partial success on a constrained industrial pick-and-place task. However, the achieved online success rate remains limited, with the best configuration reaching 42.5%, and the results should therefore be interpreted as a proof-of-concept rather than a mature industrial solution. However, the experiments also highlight significant challenges related to sim-to-real transfer, hardware limitations, motion stability and dataset quality. Differences between the evaluated models were observed in training requirements, inference behavior and task execution performance.This thesis contributes a practical implementation pipeline for humanoid VLA training and evaluation in industrial environments and provides insights into the current limitations and opportunities of humanoid robotics for industrial applications.