Project Gallery
Deep Agency - Human Guided Robotic Training for Assembly Tasks in Construction
“..., to handle products or environments with considerable uncertainty or variation, advanced force control methods are needed, such as adaptive or learning methods.”
(Apolinarska et al., 2021)
Automating robotic assembly in architectural construction presents challenges due to the unpredictable nature of construction sites and the limited dexterity of robots. Consequently, human workers often have to handle tasks that are manual and repetitive. Recent advancements in AI technologies demonstrate significant potential for enhancing the dexterity of robots. By integrating concepts from haptic teaching, deep reinforcement learning, and robotic assembly, this research examined methods to enhance robots’ performance for wood joint assembly tasks and autonomy in the construction field. Unlike the classic rule-based robotic programming methods, which require tedious hard coding for every individual customized task, the proposed workflow trains the robot through reinforcement learning, allowing the agent to explore and find solutions to assemble tasks by itself. The trained neural network is subsequently deployed to assemble lap joints of various sizes and angles that were not previously encountered by the agent. Human demonstrations are recorded on a real robot as successful experiences to improve the learning efficiency of the agent. Leveraging human knowledge, this approach empowers robots to adapt their behavior in real time to accommodate material uncertainty and deviations in timber assembly processes, granting robots greater autonomy on construction sites.
ITECH M.Sc. Thesis Project 2023: Deep Agency - Human Guided Robotic Training for Assembly Tasks in Construction
Pengfei Zhang, Sarvenaz Sardari, Selin Sevim
Thesis Advisers: Samuel Leder, Gili Ron
Thesis Supervisor: Prof. Achim Menges
Second Supervisor: Prof. Thomas Wortmann