Transformer Based Prediction of Human Motions and Contact Forces for Physical Human-Robot Interaction

As the field of robotics continues to evolve, there is a growing emphasis on achieving seamless collaboration between humans and robots. This thesis delves into the intricate dynamics of human-robot interaction. It introduces a novel approach to establish a more natural and fluid paradigm, inspired by human-to-human interactions. This approach replaces the traditional reactive model with a predictive one, where the robot doesn't merely react to human stimuli but anticipates human intentions and responds accordingly. The purpose of this thesis is to harness the power of Neural Network (NN) transformers to forecast human forces. In order to train our Neural Network we collected data from brief interactions between a human and a robotic arm, specifically a Panda arm developed by Franka Emika. In our approach, we implemented an impedance control on the robotic arm to establish a secure and harmonious collaboration. The manipulator not only ensures the safety of the interaction but also allows us to capture crucial information regarding the force exerted by the human operator to the end effector. Additionally, we captured kinematic information from the human using an Xsens motion capture suit. This tool enables the precise tracking of human movements, providing additional contextual information for our network to correctly identify the direction and the intensity of the force signal. The data collected was accurately filtered to remove external noise and given as input to custom-designed transformer architecture. The chosen network is based on GPT-2 transformer with appropriate modification to make it suitable for time series forecasting. In conclusion, the approach proposed in this study can be used to predict future forces and as a consequence to estimate the intention behind the human’s action, improving human robot collaboration.