Project Details
Projekt Print View

Active transfer learning with neural networks through human-robot interactions (TRAIN)

Subject Area Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
Human Factors, Ergonomics, Human-Machine Systems
Term from 2019 to 2025
Project identifier Deutsche Forschungsgemeinschaft (DFG) - Project number 430054590
 
Final Report Year 2025

Final Report Abstract

This project aimed to enhance autonomous robot capabilities across various environments, focusing on improving human-robot interactions (HRI). Given the complexity of manually programming diverse motor and manipulation skills, we explored Learning from Demonstration (LfD) for enabling self-improving robotic systems. LfD leverages human expertise through recorded sensor data, facilitating more intuitive and adaptive autonomous behaviors. A significant component of the project involved integrating Interactive Bayesian Optimization (IBO) to incorporate human feedback into reinforcement learning (RL) tasks, which enhanced the optimization of robotic trajectories. The Preference Expected Improvement (PEI) framework within IBO balanced user guidance and algorithmic learning, leading to more efficient skill acquisition. Additionally, we tackled the issue of posterior collapse in Variational Autoencoders (VAEs) by introducing a Contrastive Regularization (CR-VAE) approach, which greatly improved the quality of latent representations across various datasets. To understand user experiences (UX) and trust in HRI better, we examined how training data influences trust in RL agents. Results indicated that agents trained with expert human data were trusted significantly more than those with beginner data, emphasizing the importance of perceived human-like behavior in fostering user trust. Moreover, we developed a human-centered adaptive framework focusing on real-time trust assessment through physiological data. Key innovations included a sensor glove with enhanced capabilities and a Mixed Reality interface for intuitive visualization of robot actions. Findings revealed that emotional states, such as valence and arousal, directly impacted trust levels, reinforcing the necessity of adaptive autonomy to improve UX. The project demonstrated the potential of integrating UX in machine learning for reliable and intuitive interactions with robots. By enhancing user trust and HRI quality, we established a framework for effective robot deployment in collaborative environments, ensuring that robots can assist safely and efficiently in everyday tasks. The findings underscore the relevance of human guidance in optimizing robot learning processes and reliability.

Publications

 
 

Additional Information

Textvergrößerung und Kontrastanpassung