Designing Energy-Efficient Embedded Systems with CNNs for Real-Time Control in Autonomous Robotics

This research proposes an energy-aware paradigm for the implementation of deep learning using CNN for real-time control in autonomous robotics for embedded systems. The overall power efficiency of the proposed system is achieved through efficient control of power consumption, latency, and runtime while has minor degradation in terms of accuracy through a number of approaches including model pruning, quantization, and adaptive CNN scaling. Studies show that power consumption has been cut down to as low as 47.3% thereby making the optimized system's power consumption in obstacle avoidance tasks as low as 2.8 W, the base system consuming up to 5.0 W power The delay was also lowered to 47.1% in key tasks from 20 ms, thus ensuring more immediate decisions in real time activities. Also, the system up time was enhanced by 40% to perform operations including object identification and sensors integration. Accuracy stayed over 92% in all tasks; the fully optimized CNN design was 0.6% less accurate than an ideal 95.0%, with a 92.4% score. These results endorse that much energy savings and performance improvement are realizable whilst incurring moderate costs on the precision of RTAS, which makes this approach perfect for battery operated, constraint embedded systems in ‘smart’ robotics.