A Precision-Scalable Energy-Efficient DNN Accelerator for Robotic Grasping

Robotic grasping aims to perceive the environment and autonomously make decisions, necessitating low power consumption and latency. However, DNN-based grasping algorithms tend to be computationally complex. Therefore, low-bit quantization is widely used as a typical means of reducing model complexity. In this paper, we introduce a precision-scalable hardware accelerator for robotic grasping. Firstly, we propose a novel bit-compression computation method for mixed-precision, which maintains computational equivalence and further reduces complexity by 1 bit. Secondly, we develop the reconfigurable multiplication unit and low-complexity computation array. Lastly, we design an energy-efficient architecture and flexible dataflow that can flexibly support various precision of convolution and deconvolution operations. Our design is synthesized under TSMC 28nm CMOS technology. Experiments show that our work achieves a 9.23 TOPS/W energy efficiency on the mixed-precision quantized grasping algorithm. Compared to previous precision-scalable hardware accelerators, the proposed design improves up to $1.88 \times$ in energy efficiency and $1.76 \times$ in area efficiency.