Generative Neural Networks and their Application to Collaborative Robot Navigation

This dissertation explores new approaches at the intersection of machine learning and robotics, with a focus on the social navigation of robots in human-populated environments. Generative neural networks are introduced to learn human movements through observation. These models enable diverse movement predictions and forecasts of human-robot interactions, which play an essential role in the design of the navigation algorithm presented in this work. Edge cases are rare, unusual scenarios that deviate from normal conditions. Learning and evaluating these edge cases is important to ensure that a generative model can make realistic predictions. For example, humans naturally avoid collisions and plan ahead, leading to datasets where explicit collisions or evasive maneuvers are uncommon. The first part of this work focuses on developing generative models that predict the future movements of crowds while learning collision avoidance patterns. This aspect is often overlooked in the current literature but is vital for predicting the impact of a robot’s actions on humans during the planning phase. To address this problem, a new approach is proposed that does not rely on labeled data for collision avoidance. Instead, the prediction errors of a model that can lead to scenarios with potential collisions are utilized during training. These predictions are combined with a proposed collision loss function, creating a feedback loop that is minimized, effectively teaching the model how to avoid collisions. It is shown that this self-supervised learning method for collision avoidance not only improves prediction accuracy compared to the state-of-the-art but is also, to the best of knowledge, the first method to realistically learn human collision avoidance behavior without explicit labels. Building on the insights from the first part, the second part demonstrates how neural autoregressive models can split crowd movement prediction into goal and goal-conditioned trajectory models. Additionally, sampling techniques are introduced to further improve collision avoidance and increase the diversity of predictions, placing the model at the top of benchmarks. Furthermore, the goal-conditioned and collision-sensitive model enables the consideration of intended human movements and their responses to potential robot plans through hypothetical reasoning. Based on this, a Social Influence (SI) function is derived that, when minimized, promotes an interactive and less disruptive navigation strategy, which is important in shared spaces. The insights from both parts are implemented in a real robot by integrating a goal-conditioned prediction model with a model-based planning strategy. To achieve this, a hybrid model is introduced that alternates between a fast and an accurate version of the goal-conditioned prediction model. This approach fully integrates the prediction model into the planning process, enabling the robot to develop plans that not only anticipate human actions but also account for potential human-robot interactions in real-time. The goal conditioning and SI are particularly effective in balancing the expected level of interaction between the robot and humans. Additionally, the goal-conditioned model is used to select paths that resemble human-like, goal-directed behavior, guiding the robot in the desired direction. For quantitative evaluation, a benchmark environment based on real-world scenarios is implemented, demonstrating the potential of the proposed method. Furthermore, experiments conducted on a robotic platform confirmed the effectiveness of the methodology in real-world scenarios. The dissertation concludes by highlighting the importance of feedback loops in model training, the integration of prediction and planning, and the effectiveness of the model-based planning strategy in real-world scenarios. It also emphasizes the need to plan for collaboration with humans by viewing them as active participants rather than just obstacles. Further research is recommended to refine robot navigation and explore how machine learning can further enhance human-robot interactions.