ResNet-18 based multi-task visual inference and adaptive control for an edge-deployed autonomous robot

Area of Science:

• Robotics
• Artificial Intelligence
• Edge Computing

Background:

• Industrial robots are often too complex and expensive for SMEs.
• Existing solutions may rely heavily on external computing resources, limiting autonomy.

Purpose of the Study:

• To develop a cost-effective, autonomous logistics robot for SMEs.
• To integrate adaptive control, visual perception, and mechanical handling on an edge platform.

Main Methods:

• Utilized an NVIDIA Jetson Nano with a modified ResNet-18 model for concurrent task execution.
• Implemented a lightweight rack-and-pinion mechanism for payload lifting (up to 2 kg).
• Integrated object-handling zone recognition, obstacle detection, and path tracking.

Main Results:

• Achieved 92% path tracking accuracy, 88% obstacle avoidance success, and 90% object handling success.
• Demonstrated a maximum perception-to-action latency of 150 ms.
• Maintained stable operation for up to 3 hours on a single charge.

Conclusions:

• The developed robot offers a practical, low-power, standalone automation solution for SMEs.
• This integrated system addresses limitations of complexity, cost, and external dependencies in current industrial robotics.
• The robot shows significant potential for enhancing logistics and operational efficiency in SMEs.