Learning manipulation skills for the Franka FR3 using behavior cloning: from simulation to real-robot deployment

Abstract

This thesis presents a behavior cloning pipeline of robotic manipulation on the Franka FR3, including simulation validation, real-robot data collection, policy train-ing, and the deployment on hardware. The superiority of temporally-aware architectures is validated by a simulation study of BC-RNN and Diffusion Policy on the Robomimic benchmark, where Diffusion Policy has 98-100% success in Lift and Pick-PlaceCan tasks compared to 78-92% in BC-RNN. Action Chunking with Transformers (ACT) is chosen to be deployed on real robots due to its capability to work with low-data regimes and single-pass inference. Three demonstration datasets were gathered through joystick teleoperation through an iterative process with each dataset being redesigned by the lessons of the previous deployment. These datasets were trained on four ACT policies with the LeRobot framework in 100,000 steps, and the training loss dropped to 0.038. The policies were executed on the physical robot through a ROS 2 network through CRISP controllers and CycloneDDS, and nine software integration issues had to be solved. The system improved in four deployment cycles, starting with no grasp attempts generated in the first deployment, and ending with the full pick-and-place task structure - approach, grasp attempt, lift, transport, and placement - generated in the last deployment. The main gap was the precision of spatial localization during the grasp phase, and reliable task completion was not achieved. This thesis not only provides a completely documented iterative pipeline but also pinpoints camera location, demonstration action allocation, and deployment time synchronization as the most important practical considerations in real-robot BC performance.