Architectural Design and Analysis of an Autonomous Robot Based on Edge AI, Computer Vision, and Reinforcement Learning for Intelligent Navigation in Smart Environments
Architectural Design and Analysis of an Autonomous Robot Based on Edge AI, Computer Vision, and Reinforcement Learning for Intelligent Navigation in Smart Environments
Arash Azizpor1
1) Undergraduate Student of Electrical Engineering (Electronics Concentration), National Skills University, Kermanshah, Iran
IOI Code
:
ACFD-ACZGBZ
Publication :
(https://3nbececonf.ir/fa/panel/papers)
Abstract :
Autonomous mobile robots deployed in smart
environments such as hospitals, warehouses, universities, and
smart buildings must comprehend crowded, dynamic scenes, plan
safe movement, and react within tens of milliseconds, all while
operating under strict constraints regarding power, cost, and
network connectivity. Cloud-based designs suffer from latency,
bandwidth costs, and privacy concerns, whereas purely classical
navigation stacks struggle in spaces densely populated by
humans. This paper addresses the architectural design and
analysis of a five-layer autonomous robot that integrates
computer vision, Simultaneous Localization and Mapping
(SLAM), and a Deep Reinforcement Learning (DRL) based local
navigation policy onto an Edge AI platform, coordinated by ROS
2 and Nav2. We formulate models for kinematics, perception,
mapping, and decision-making (differential-drive kinematics,
back-projected depth imaging, log-odds occupancy, Markov
Decision Process, Q-learning and PPO objectives, and Control
Barrier Functions); derive a latency budget and a stoppingdistance constraint that limits the robot's permissible speed; and
describe a simulation-to-edge deployment pipeline based on
domain randomization, INT8 quantization, and TensorRT. To
analyze the behavioral trade-offs involved in combining a global
planner with a learned local policy, a study using a dynamic gridworld environment (featuring 500 test episodes, five moving
obstacles, and randomized maps with wall and gap traps)
compares reactive strategies, A* with replanning, pure
reinforcement learning, and a hybrid approach. The hybrid agent
achieved an 80.2% success rate with a 3.2% collision rate; in
contrast, A* with replanning yielded rates of 89.2% and 10.8%
respectively, while pure reinforcement learning achieved a 59.4%
success rate. The results demonstrate that global guidance is
essential for learned local policies and that such policies can
significantly reduce collision rates, albeit at the cost of path
efficiency. This article concludes by outlining limitations and
presenting a roadmap for future research, which includes guiding
and deploying foundation models, safe reinforcement learning,
and federated learning at the network edge.
Keywords : autonomous mobile robot edge ai computer vision deep reinforcement learning intelligent navigation ros 2 smart environments model quantization