|
|
|
|
|
Deep Timber, ETH Zurich, 2018-2020
|
Reinforcement Learning for Robotic Assembly of Timber Joints
|
This research project investigates the application of machine learning to facilitate architectural construction of timber structures using industrial robots. We employ Deep Reinforcement Learning to teach a robot to assemble lap jointed timber beams, i.e. to insert one notched beam into another, based on contact forces sensed by the robot. In comparison with traditional robot programming, the learning approach enables real-time adaptation to inaccuracies and initial misalignment.
Specifically, we use force-torque sensor data and end-effector poses (observations) to learn sequences of robot movements (actions) that lead to a successful completion of the assembly task. Using an adapted policy gradient method, we leverage human demonstrations and train entirely in a physics simulation environment. We show that these trained policies can be deployed on real robots, thus successfully bridging the “reality gap”. We also demonstrate that our policies can generalize to a fair range of situations not encountered in training, for example, to successfully assemble joints which have a slightly different shape, size or initial position. In a broader picture, this translates into an ability of the robotic system to respond to uncertainties and inaccuracies in the real world.
Publications:
Apolinarska, Aleksandra Anna, Matteo Pacher, Hui Li, Nicholas Cote, Rafael Pastrana, Fabio Gramazio, and Matthias Kohler. Robotic Assembly of Timber Joints Using Reinforcement Learning. Automation in Construction 125 (May 2021): 103569. DOI: 10.1016/j.autcon.2021.103569. PDF
|
|
|
|
|
Credits:
|
Gramazio Kohler Research, ETH Zurich
In cooperation with: Autodesk Research, Autodesk Inc., San Francisco (Dr. Hui Li, Nicolas Cote, Erin Bradner) Collaborators: Dr. Aleksandra Anna Apolinarska, Matteo Pacher, Lukas Stadelmann, Rafael Pastrana, Hannes Mayer
|
|
De |