Autonomous Driving Systems (ADS) are safety-critical applications designed to enhance vehicle autonomy and general road safety. Perception, a critical stage of the pipeline, detects and classifies objects and environmental conditions. This stage is also highly susceptible to faults such as transient and permanent bit flips, which compromise the entire ADS pipeline. Conventional Triple Modular Redundancy (TMR) and Error Correcting Code (ECC) solutions have well-known limitations in satisfying strict ADS requirements. TMR primarily masks the faults by redundancy but does not include error correction or end-to-end fault management capabilities. ECC works well for transient faults like one or two bit flips but fails to handle multi-bit error scenarios.
GW Researchers have developed a novel chiplet-based self-healing hardware architecture tailored to address the vulnerabilities of the perception stage in the Autonomous Driving Systems (ADS) pipeline, particularly in image processing and object detection. The proposed mechanism ensures robust CNN operations with minimal performance overheads, even under stringent real-time constraints, by autonomously detecting, diagnosing, and recovering from faults. Leveraging FPGA-based technology, proposed architecture combines high performance, scalability, and adaptability, meeting ADS systems’ reliability and fault-tolerance demands.
Figure: Self-Healing Architecture
Advantages:
- Detects, corrects, and resolves faults efficiently
- Efficient correction for both data and logic faults
- Achieving low latency of 0.95092 ms, and power consumption of 5.8 W
Applications:
- Advanced Driver-Assistance Systems (ADAS)
- Perception stage of the ADS pipeline
