[MA] Out-of-Distribution Detection of Large Neural Networks for Safe and Efficient Human-Robot Collaboration

14.03.2024, Abschlussarbeiten, Bachelor- und Masterarbeiten

We are on the verge of a technological revolution, where robots are becoming increasingly integrated into our everyday lives. New robots must interact in highly dynamic environments and guarantee safety for themselves and others. For this, measuring the environment as accurately as possible is essential. Cameras are cheap and flexible sensors for most robotic applications with the potential to measure the entire environment. Deep neural networks have been demonstrated to be the most sufficient solution for extracting information from images. Unfortunately, we cannot give strong bounds on the disturbances that act on the output of a deep neural network. For safety-critical perception like human detection and prediction, we often cannot rely on deep neural networks, and we restrain ourselves from using them altogether. To alleviate this, we propose a framework for a safe robotic stack that integrates deep neural network predictions into formal methods to give strong safety guarantees. We aim to deploy this framework in human-robot collaboration settings where a manipulation robot interacts with humans. Deep learning underlies two sources of uncertainties: aleatoric uncertainties are due to inherent randomness and imprecision in the learning process, whereas epistemic uncertainties mainly stem from states unseen during training time. Predicting aleatoric uncertainties became a common standard in most deep learning domains. Recent advances in out-of-distribution detection predict epistemic uncertainties, allowing us to predict when the model output is reliable. As shown in previous works, these methods can then be used to guarantee the safety of autonomous robots by falling back to a failsafe trajectory if an out-of-distribution state is detected. Our proposed safety framework utilizes set-based reachability analysis [3, 6]. Its core idea is to predict all possible future states of the robot and the environment. We then validate the safety of this set of possible future states. If a subset of future states is recognized as potentially unsafe, we fall back to a failsafe trajectory, bringing the robot to a safe state. We propose using neural network prediction to measure parts of the robotic stack. The predicted aleatoric uncertainty will be used to determine the size of our reachable sets. We can discard false measurements by detecting out-of-distribution inputs of the neural network. Our set-based reachability analysis propagates the uncertainties in the system forward in time, regardless of the last measurement time. We can, therefore, guarantee continuous safety regardless of out-of-distribution states. We want to show that safe human-robot collaboration is possible based on visual inputs alone. For this, we propose two prediction networks, one for measuring the current human pose and one for predicting future human motion. Both networks will be checked for aleatoric and epistemic uncertainties to ensure a prediction within known bounds. By integrating this prediction into our reachability analysis, we can shrink the human reachable sets, making the robot control less restrictive while maintaining high robustness. 'Full thesis proposal'

Kontakt: jakob.thumm@tum.de

Mehr Information

https://www.ce.cit.tum.de/fileadmin/w00cgn/cps/_my_direct_uploads/Thumm_2024_MA_Proposal_Out-of-Distribution_Detection_for_Safe_and_Efficient_Human-Robot_Collaboration.pdf