Research Needs

• Advanced Reasoning and Learning
  • Ability to reason causally about real-world actions, including the vehicle itself, other vehicles, and pedestrians
  • Learning approaches from shared observations from a network of connected, fully autonomous agents
  • Use of contextual explanatory models to justify decisions and to enable improved performance of the system
  • Modification of the driving style based on the identity and state of the passenger
  • Intuitive, intelligent user interfaces for modifying system operation
• Robust security of systems, components, and networks
  • Computing architectures for improved machine learning and decision making algorithms
  • Decision frameworks and algorithms for reliable sensor fusion and safe and secure actuation
  • Efficient and robust overall system design and architecture
  • Methods to reduce computation for automotive machine vision neural networks
  • Computation optimized radar-based pedestrian detection
  • Sensor Fusion and Perception
• Multimodal fusion to substantially improve performance over conventional object-level fusion, where each modality contributes an independent detection result
  • Use of intermediate sensor responses (e.g. extracted features) to dynamically tune, align and calibrate multimodal systems
  • Multi-sensor and multi-modal context-aware sensing that integrates sensors and processing to provide significantly enhanced sensor output
  • Robust environmental models that outperform those available from conventional sensor systems by more accurately capturing, modeling, and tracking complex, dynamic scenarios in challenging conditions
  • Approaches for either outlier rejection, risk based decision processes, or compensation for known sensor weaknesses
• Communications
  • V2X communication, sensor communications and advanced security protocols
  • Algorithms for self-optimization of many vehicles and for robust collaboration among cars
  • Advanced crypto algorithm development for performance and attack resilience
  • Advanced authentication and verification methodologies
• High Dependability
  • Self-adaptive systems to improve efficiency, including continuous self-optimization
  • Anomaly detection to promote functional safety
  • Self-healing, on-line system reconfiguration to handle hardware, sensor and software failures
  • Dependable operations using a bounding envelope for constraining the range of allowable AI actions
  • Safety/security/dependability co-design to ensure robust and safe operation under all conditions