Remote Object Pose Estimation for Agile Grasping: Leveraging Cloud Computing through Wireless Communication

Abstract

As industry transitions toward Industry 4.0, the demand for agile robotic systems capable of vision-guided manipulation is rapidly increasing. However, the computational limitations of onboard hardware make it challenging to support advanced perception pipelines, particularly those based on deep learning. Offloading perception to the cloud presents a promising alternative but introduces latency and reliability challenges that can compromise the real-time performance required for closed-loop robotic control. This thesis presents a robotic grasping system capable of agile, 6D pose-aware manipulation of moving objects by offloading perception to a remote inference server. RGB-D data is continuously streamed over a wireless link to the server, where a deep learning model estimates the object's 6D pose. The estimated pose is then sent back to the robot, which uses it to generate a trajectory for executing the grasp. The system was evaluated on a conveyor-based pick-and-place task under four different wireless network types: Wi-Fi at 60 GHz, Wi-Fi 5 at 5 GHz, 5G NSA at 24 GHz, and 5G NSA at 3.5 GHz. A total of 392 trials were conducted to analyze grasping success rates and the impact of network latency and reliability on performance. The results demonstrate the feasibility of performing agile, closed-loop robotic grasping with cloud-offloaded 6D pose estimation over wireless networks. They also reveal limitations of current wireless infrastructure and deep learning models. The findings suggest that lower-latency, more reliable communication, along with more intelligent local control strategies and faster, generalizing models, are required for production deployment.