Session: 15-03-01: Machine Learning and Deep Learning in Safety, Reliability, and Maintenance I

Paper Number: 166493

Towards Reliability Assessment of AI-Controlled Robotic Systems Through Deep Learning-Based Skill Detection 

The integration of AI-controlled robotic systems into industrial environments has significantly advanced automation, enabling robots to perform complex tasks alongside human operators. However, this integration also presents critical challenges, particularly in ensuring the safety and reliability of these systems. Traditional reliability assessment techniques often struggle to adequately evaluate the components of these black-box systems. Therefore, a dynamic reliability assessment approach is essential to account for the variability and complexity of AI-controlled robotic systems.

In [1] we proposed the concept for a dynamic reliability assessment framework for AI-controlled robotic systems. It consists of five key methods: (i) Data Logging – Capturing extensive robotic simulation data to document system behavior; (ii) Skill Detection  – Employing deep learning techniques to automatically identify and categorize robotic skills such as pick and place; (iii) Behavioral Analysis – Developing a detailed behavioral profile to assess performance under varying conditions; (iv) Reliability Model Generation – Computing failure probabilities of robotic hardware components to create advanced hybrid reliability models; and (v) Reliability Model Solvers – Numerically evaluating hybrid reliability models to quantify potential risks. This structured approach enables a more precise and flexible evaluation of AI-controlled robotic systems, addressing the limitations of conventional reliability assessment methods.

In this paper, we introduce the first two key methods of this framework: Data Logging and Skill Detection. These components are fundamental to the dynamic reliability assessment process, as skill detection provides a way to interpret black-box AI policies, while data logging captures structured insights into robotic system performance.

We created two time-series datasets using the Franka Emika Panda in simulation, covering both basic and complex skills. The primary distinction between these datasets is that the same skills were performed in a more complex manner by the robotic manipulator, requiring the coordination of additional joints. These datasets provided a robust foundation for training and testing.

We implemented four deep learning architectures specifically designed for skill detection: CNN, LSTM, CNN-LSTM, and Transformer. Performance comparisons revealed that the CNN-LSTM model outperformed others, achieving an accuracy of 99.2% for basic skills and 92.61% for complex skills in simulation. 

The model was further tested across different robotic manipulators, including the Jaco, UR5, OpenMANIPULATOR-X, and Franka Emika for Cable Routing. The Jaco, UR5, and Franka Emika for Cable Routing are open-source datasets from the OpenX Embodiment dataset, while the OpenMANIPULATOR-X dataset was collected in our lab. This cross-platform evaluation demonstrated the robustness of the model across various embodiments. Notably, models trained from scratch consistently outperformed those utilizing transfer learning, particularly when applied to real-world data. 

Our skill detector can be easily extended beyond robotic manipulators to include Autonomous Mobile Robots by adapting the model to recognize navigation, manipulation, driving, and interaction tasks. However, incorporating these new skills requires retraining the model on domain-specific datasets to ensure accurate classification and adaptability. This retraining process allows the network to learn task-specific features, improving its generalization across different robotic platforms.

This work lays the groundwork for enhancing the safety and reliability of AI-controlled robotic systems through advanced skill detection techniques. By leveraging deep learning for dynamic reliability assessment, this research contributes to the development of more dependable and risk-aware robotic automation solutions.

References

[1] Grimmeisen, Philipp, Friedrich Sautter, and Andrey Morozov. "Concept: Dynamic risk assessment for ai-controlled robotic systems." arXiv preprint arXiv:2401.14147 (2024).

Presenting Author: Philipp Grimmeisen University of Stuttgart

Presenting Author Biography: Philipp Grimmeisen is a PhD candidate at the University of Stuttgart, specializing in the reliability and risk assessment of robotic systems. His research focuses on evaluating the reliability of robots controlled by black-box policies, leveraging techniques such as simulation, fault injection, skill detection, automated generation of reliability models and probabilistic modeling to assess and improve their reliability and safety.

Authors:

Rucha Golwalkar University of Lübeck
Philipp Grimmeisen University of Stuttgart
Friedrich Sautter University of Stuttgart
Andrey Morozov University of Stuttgart