Robotic assembly of timber joints using reinforcement learning

In architectural construction, automated robotic assembly is challenging due to occurring tolerances, small series production and complex contact situations, especially in of elements with form-closure such as timber structures integral joints. This paper proposes apply Reinforcement Learning control robot movements contact-rich tolerance-prone tasks presents the first successful demonstration this approach context construction. Exemplified by lap joints for custom frames, are guided force/torque pose data insert a element its mating counterpart(s). Using an adapted Ape-X DDPG algorithm, policy trained entirely simulation successfully deployed reality. The experiments show can also generalize situations real world not seen training, tolerances shape variations. caters uncertainties construction processes facilitates fabrication differentiated, customized designs.