3.1. Proposed work

The paper proposes a robotic architecture based on the IoT that will make it possible to overcome the challenges of real-time navigation that is inherent to the mobility of smart campuses and cities by implementing smart perception, decision-making, and control into the same robotic system.

In contrast to the traditional GPS-oriented or rule-of-thumb solutions that are unable to work in semi-structured spaces (corridors, intersections, or crowded pedestrian areas) and fail to work when they have to, the system is hypothesized to provide real-time flexibility with onboard smartness. The architecture of the invention is founded on three fundamental layers, including a perception layer with IoT sensors (ultrasonic, LiDAR, IMU, and GPS) to sense, a hierarchical reinforcement learning (HRL)-based decision layer to decide, and a low-latency control layer to control the robot movement. A Kalman filter is an extended Kalman filter that will do sensor fusion to give it localization strength and resilience as far as obstacle detection is involved. The fact that these IoT sensors are low cost and compact is also so that the system can be scaled and able to run on resource-intensive public fixtures like school campuses that expensive 3D-LiDAR-based SLAM systems require. The HRL is the structure of a decision engine where a high-level policy selects the high-level strategy navigation goals and a low-level policy selects continuous control actions such as path following, rotating, and avoiding collisions. The RL agent also trains the PPO in a ROS-Gazebo simulation environment, where it is free to run on real-world university campus maps and traffic routes. PPO is compared to naive Q-learning or DDPG, as it can operate in a continuous action space and converges reliably in a large range of diverse environments. One of the most interesting developments is the multi-objective rewarding functionality, which optimizes travel time, energy usage, and safety at the same time, unlike the traditional RL methods that are more likely to aim at and target one objective, such as the shortest distance. Furthermore, it is implemented on an edge-AI platform (NVIDIA Jetson Nano) in real time, it does not rely on the cloud, and it allows autonomous decision-making in the operating scenario. Campus robotics, in particular, cannot have network dependability. The system is analyzed for the proposed system in three scenarios: (i) dynamic interference of obstructions in urban traffic, (ii) GPS-deprived campus corridors and multiple-turn routing, and (iii) noisy environment with low sensor exposure. Comparative simulations reveal that the proposed system considerably outperforms conventional A* routing, reactive rule-based planners, and flat RL agents on navigation success rate, path efficiency, and obstacle avoidance. In conclusion, this paper adds a new, scalable, and deployable field robot platform that combines IoT-based environmental sensing with deep RL in a combined package that addresses directly both the smart campus and the smart mobility navigation challenge. With the provision of modularity, edge-processing autonomy, and multi-context flexibility, the system differs from other standalone navigation solutions that focus on autonomous robotics, field testing, and novel algorithms.

3.2. Workflow of the proposed framework

The process flow of the suggested framework is a linear process with real-time processing and adaptable navigation. The functioning of the suggested IoT-based adaptive robot system depends on a module-based real-time processing pipeline to offer smart and context-aware navigation along intelligent campus pathways and urban transport routes. The process starts from data gathering via a distributed array of IoT sensors on the mobile robot supplemented by GPS for outdoor positioning, LiDAR for 3D spatial awareness, and cameras for vision insight. For example, GPS logs are refreshed every 1 s, LiDAR scans at a 10 Hz rate, and camera modules with a detection accuracy of more than 95% detect traffic signs and lane markings during the day. Raw data streams are filtered and preprocessed with a Kalman filter-based data fusion module to provide homogeneous, noise-free spatial perception. For instance, the use of LiDAR in conjunction with GPS reduces the positional accuracy to 10 cm, which is essential for motion control in narrow spaces. The visual perception module utilizes a CNN at 30 fps and identifies pedestrians, automobiles, and static road features with an average accuracy rate of 98%. These identifications produce semantic labels and spatial flags for pedestrian obstacle priority. For example, the system initiates an avoidance protocol when the system detects a cyclist at 5 m. The path planning module is controlled by RL, with the agent that dynamically learns in order to weigh potential paths based on real-time traffic congestion, destination distance, and energy cost. The system is capable of computing optimal paths in less than 2 s to ensure the highest degree of reactivity, and in a multi-choice form, the RL reward structure will select a path according to the two objectives: the safety criterion and the efficiency criterion. Moreover, the recurrent neural network (RNN)-based predictor is used to forecast previous and current traffic to forecast future traffic congestion with 90% accuracy 15 minutes in the future. This enables the robot to undergo proactive yielding by working toward a region of future congestion (bottleneck at an intersection). Trajectory planners (e.g., A + or rapidly exploring random trees (RRT)) are developed by combining the camera with the LiDAR stream in real-time collision avoidance. The trajectory planners produce actuation latency-insensitive smooth collision-free trajectories required to allow safe operation in congested environments with less than 200 ms. It will also have adaptive behavior with smart decision loops in the system. When sensor readings (e.g., the estimator is operating based on the inertial estimates due to the loss of the GPS signal) are not available, the data readiness check is used to detect it. The path check optimization will substitute the existing path with new cost functions. The safety check is used to make sure that an unsafe pathway is not created. All the feedback mechanisms collectively constitute a closed-loop system that offers continuous adaptability, insensitivity to uncertainty, and security guarantees. This modular workflow ensures that at each stage, from perception to control, data is transformed into actionable intelligence, enabling the robot to navigate autonomously in rapidly changing environments. Table II summarizes the structured pipeline with its associated inputs, processing modules, and output actions, offering a clear view of how IoT, deep learning, and robotic navigation are cohesively integrated to achieve real-time, adaptive operation.

The adaptive and efficient system will handle ever-changing urban traffic circumstances that require real-time data processing and intelligent decision-making guidance. The workflow steps will have utilize defined input and processing at each step to result in an actionable output created from those definitional properties of the defined data and maintain both adaptive system potential and efficiency throughout the series of changing traffic dynamic conditions. A complete overview of the entire operational pipeline of the proposed system is illustrated in Fig. 1, which shows a structured mathematical flow for adaptive navigation. The proposed system will commence its workflow by receiving sensory data, obtained from multimodal sensors, which may require Kalman filter fusion techniques to enable more accurate analysis and localization of the self-driving vehicle. The system then performs object classification through CNN-based visual perception and anticipates congestion using an RNN model. RL via PPO governs the path planning process by optimizing a multi-objective reward function. Collision-free trajectory generation is executed using A* and RRT algorithms, followed by a final actuation step. Real-time safety checks and fallback mechanisms are integrated to ensure robustness in sensor failure or unpredictable traffic behavior. This modular structure ensures that each decision is both data-driven and context-aware.

Table II.

Workflow of the proposed framework for AI-powered robotic navigation. This table summarizes the data used, processing techniques applied, and output of each step of the workflow, integrating IoT, deep learning, and robotic navigation for real-time management of urban traffic.

Figure 1.

Workflow diagram of the proposed adaptive robotic navigation system integrating IoT sensing, Kalman-based data fusion, deep visual perception, traffic prediction, reinforcement learning-based decision-making, and optimal trajectory execution. The flowchart outlines the sequential operation of the system, key processing modules, embedded mathematical formulations, and real-time decision feedback loops used for safe and efficient autonomous mobility in dynamic environments.

3.3. Data generation and optimal path selection

Full data-driven route optimization is a vital part of the aforementioned adaptive navigation framework in Fig. 1, outlining the end-to-end system workflow. In this framework, raw sensory data collection through IoT-infused hardware such as GPS, LiDAR, IMU, and RGB cameras initiates the process. The data streams are combined through an extended Kalman filter to generate accurate localization estimates with error minimization from each sensor. Real-time visual information is processed using a CNN to identify and classify objects such as cars, pedestrians, and traffic signs. In the background, past and real-time traffic information is processed using an RNN model to forecast congestion in the future in front to enable anticipatory decision-making. Robot trajectories are constructed from a PPO RL policy that optimally co-controls a multi-objective reasonable/desired reward function for time and energy. The trajectories of a robot are computed and implemented using conventional motion planning techniques (i.e., A* and RRT) to generate dynamically feasible and collision-free motion. The actuation is executed with closed-loop feedback controller laws that take into account safety, sensor invariance, and performance. While Fig. 1 only shows a flowchart of the overall workflow, the design of the data acquisition, data fusion, predictions, and decisions modules to calculate optimal and adaptive robot trajectories in real-time is described in detail.

3.3.2. Path optimization

The process of selecting the best route within the provided context is governed by an RL system that produces direction-of-motion decisions dynamically as a function of the sensed real-time environment. Unlike traditional navigation strategies employing static maps or deterministic planners like Dijkstra or A*, the system here employs PPO to learn adaptive policies for navigating uncertain, obstacle-laden, and dynamic environments. The RL agent responds to an aggregated state vector containing LiDAR-based 3D obstacle information, camera-detected semantic label information, and Kalman filter-localization estimates of GPS and IMU data. The agent takes an action in its environment via the learned π(at∣st;θ), where at is the selected action, st is the state, and θ are policy parameters. The originality in this research lies in the fact that a multi-objective reward function is learned, enabling the robot to exchange performance metrics other than mere arrival at the goal. The reward R at time step t is utilized in Eq. (1):

(1) \begin{equation}Rt=w1\left(-Tt\right)+w2\left(-Et\right)+w3\left(-Ct\right)\end{equation}

where Tt refers to travel time to the destination, Et refers to power consumption, Ct refers to the penalty for collision in distance from obstacles, and w1, w2, and w3 refer to scaling factors to assign relative priority among speed, efficiency, and safety. This formulation makes the robot learn to minimize travel time and power consumption and also to avoid dangerous maneuvers near obstacles. The RL policy is learned with PPO, exhibiting sample efficiency and convergent stability on continuous control tasks. Once trained, the policy is deployed in a real-time environment using lightweight edge computing (NVIDIA Jetson Nano), enabling onboard execution with low latency. The control structure is hierarchical: the high-level PPO agent selects optimal waypoints, while a low-level planner (e.g., A* or RRT) refines those into collision-free trajectories. Safety and optimality checks are embedded in the loop to reevaluate decisions as environmental conditions evolve. The optimal path is determined using the A* algorithm or RL, defined by the cos function as using Eq. (2):

Cost Function:

(2) \begin{equation}J=\sum _{i=1}^{n}(w1.di+w2.ti+w3.ei) (2)\end{equation}

Where:

d_i : Distance to the destination for each segment.

t_i : Estimated time to traverse each segment.

e_i : Energy consumed while traversing each segment.

w₁, w₂, w₃: Weights assigned to prioritize distance, time, or energy.

Application: For a route divided into n segments, the algorithm calculates the cumulative cost J and selects the route with the lowest cost. This ensures shorter travel times, reduced energy consumption, and avoidance of congested routes.

Sample Calculation:

• • Path divided into n = 5 segments.

• • Parameters: d = [200,300,150,250,400] (meters), t= [30,45,20,35,60] (seconds), e = [500,750,300,600,1000] (joules).

• • Weights: w1 = 0.5, w2 = 0.3, w3 = 0.2.

Step-by-Step Computation:

1. 1. Segment 1: J1 = (0.5⋅200) + (0.3⋅30) + (0.2⋅500) = 100 + 9 + 100 = 209J.

2. 2. Segment 2: J2 = (0.5⋅300) + (0.3⋅45) + (0.2⋅750) = 150 + 13.5 + 150 = 313.

3. 3. Repeat for all segments.

4. 4. Total Cost:

J = J1 + J2 + J3 + J4 + J5 = 209 + 313.5 + 141 + 255.5 + 418 = 1337J

Obstacle Avoidance: Avoidance is modeled using dynamic thresholds; see Eq. (3):

Safe Distance:

(3) \begin{equation}D_{safe}\gt \sqrt{\left(x_{obs}-x_{\textit{robot}}\right)^{2}+\left(Y_{obs}-Y_{\textit{robot}}\right)^{2}}\end{equation}

Where:

D_safe: Minimum safe distance.

x_obs, y_obs : Coordinates of the obstacle.

x_robot, y_robot : Coordinates of the robotic vehicle.

Application: The model continuously computes the distance to nearby objects. If the distance is less than the safe threshold, avoidance mechanisms (like rerouting or deceleration) are triggered.

Sample Calculation:

• • Robot position: (2,3).

• • Obstacle position: (5,7).

• • Safe distance: Dsafe = 5 m.

Compute the Euclidean distance:

D= (5−2)²+(7−3)² = 3² + 4² = 9 + 16 = 5

If D ≥ D_safe, the system initiates avoidance measures.

Traffic Prediction: RNN-based models predict traffic using, as in Eq. (4):

(4) \begin{equation}T_{next}=f\left(T_{past},T_{\textit{current}}\right)\end{equation}

Where:

T_next: Predicted traffic condition

T_past, _Tcurrent: Historical and real-time traffic data.

Application: By forecasting congestion, the system proactively adjusts the route to avoid bottlenecks, ensuring smoother navigation.

Sample Calculation:

• • T_past= [Reference Mondal and Rehena20, Reference Singh, Kaunert, Lal, Arora and Jermsittiparsert24, Reference Gopal, Gupta, Sharma, Kaushal, Joshi and Sharma28] vehicles/minute.

• • T_current = 35 vehicles/minute.

Using an RNN with learned weights, the model predicts $T_{next}$ = 38 vehicles/minute. This prediction triggers a proactive rerouting decision.

The design and execution logic of the RL-based optimal path selection module are illustrated in Fig. 2, which presents a composite view of the system’s internal architecture, learned reward dynamics, control policy behavior, and overall functional flow. These subfigures highlight the hierarchical integration of sensor-driven state estimation, multi-objective reward modeling, and trajectory planning. The reward heatmap and policy map further confirm the effectiveness of the RL agent in capturing optimal behaviors under diverse environmental conditions, thereby validating the robustness and adaptability of the proposed decision system.

Figure 2.

Comprehensive visualization of the proposed optimal path selection framework. The collage includes (a) a real-time RL-based workflow diagram showing sensor fusion, reward computation, and trajectory generation; (b) a heatmap visualization of cumulative reward across different state-action pairs indicating policy learning convergence; (c) a system-level flowchart integrating perception, prediction, planning, and actuation modules; and (d) a reinforcement learning policy map visualizing the navigational decisions in structured environments. Together, these visualizations reflect the functional depth, mathematical rigor, and real-time applicability of the proposed navigation framework.

3.4. Simulation setup and implementation

The simulation setup includes three primary scenario settings: urban traffic simulation, handling obstacles, and weather variability. Each of these settings intends to mimic real-life phenomena, thus creating an efficient testbed for validation purposes.

Feedback Loops and Metrics:

• • Traffic Resilience : The feed of real-time traffic density into the RL model assists in dynamically adjusting the reward function to make optimal decisions.

• • Obstacle Avoidance Performance : Collision detection metrics for true positives and false positives are logged to assess obstacle handling efficiency.

• • Robustness to Weather Conditions: Testing sensor reliability in varying conditions of visibility; performance metrics include success rates in navigation and average deviation from the optimal Please check sentence “Testing sensor reliability in…” for clarity.path.

Validation through 100 test cases demonstrates an overall success rate of 95% for tasks related to navigation. The feedback loops ensure continuous improvement in the framework by adapting well to dynamic conditions in an urban environment. The simulation setup and the implementation framework for the proposed AI-powered navigation system, which integrates hardware components, software tools, and test scenarios, are presented in Fig. 3. Such a framework ensures testing and validation under diverse conditions with real-world realism.

Figure 3.

Simulation setup and implementation framework. This figure presents the structural flow of the proposed simulation framework, integrating hardware components, software tools, and test scenarios for the development and validation of the AI-powered robotic navigation system.

The proposed simulation setup and the implementation framework are integrating the hardware components, software tools, and test scenarios in an organized manner as depicted by the diagram. The collection and processing of environmental data are done by hardware components like NVIDIA Jetson Nano/GPU, LiDAR sensors, GPS modules, cameras, and robotic navigation platforms. These then feed into the software tools such as TensorFlow, PyTorch, ROS, Gazebo Simulator, and MATLAB, which handle AI model training, real-time robotic operations, and virtual environment simulations.

The framework is tested with three primary scenarios: urban traffic simulation for testing navigation in variable traffic densities, obstacle handling to ensure collision-free navigation, and weather variations for validating system robustness under varying conditions. Within the Urban Traffic Simulation, there were three different lanes built to simulate different traffic volumes: light, moderate, and heavy. Lane 1 was light traffic in which vehicles were placed thinly, Lane 2 simulates moderate traffic similar to what is experienced in general urban congestion, and Lane 3 simulates heavy traffic with dense vehicle placements. The system is tested for traffic conditions, which it would then determine the best route that has minimal congestion while it can navigate through with maximum efficiency. Figure 4 highlights the optimal route through the least congested lane in green.

Figure 4.

Urban traffic simulation. The system effectively identifies and selects the optimal route in light traffic conditions.

The Obstacle Handling scenario aimed to validate the system’s ability to navigate safely with respect to static and dynamic obstacles. The static obstacles referred to the parked vehicles or barriers, while the dynamic obstacles corresponded to moving entities, such as pedestrians or cyclists. The system could successfully identify the obstacles and make collision-free path plans, which it could adapt to changes in the environment. The effectiveness of the proposed obstacle avoidance mechanism is illustrated in Fig. 5, which showcases the robot navigating in a test environment populated with both static and dynamic obstacles. The top view reveals how the robot computes a safe trajectory that curves around obstructions, while the front view demonstrates spatial awareness in vertical positioning and path height clearance. The color-coded legend enhances interpretability, showing optimal and rejected trajectories alongside the real-time obstacle map. This figure validates the policy’s collision-free decision-making in constrained environments, reinforcing the robustness of the RL framework used for safe robotic navigation.

Figure 5.

Obstacle handling highlights the system’s collision-free navigation path in the presence of static and dynamic obstacles.

In the Weather Variations scenario, the system’s performance under challenging environments was tested with adverse weather conditions of fog or low visibility. Fog was represented using a semi-transparent overlay, and gradients and visual effects were used to introduce rain and low-light conditions. The system maintained safe navigation by finding a clear path with the minimum obstruction in terms of vision, as shown in Fig. 6. That scenario checked the reliability of sensors and the robustness of the decision-making procedure under various weather conditions.

Figure 6.

Weather variations. The figure demonstrates the system’s ability to handle foggy conditions and maintain visibility for route optimization.

The data flow process starts with the hardware, where real-time inputs are being collected, which are processed by the software tools and validated through test scenarios. This setup provides a scalable and modular approach for validating the efficiency, adaptability, and safety of the framework under realistic and diverse operational conditions.

The structural overview of the proposed simulation setup and implementation framework is presented in Fig. 7. The diagram below showcases modular hardware and software layer integration, which includes real-time data acquisition from sensors such as GPS, LiDAR, and cameras, and on-device processing with an embedded system (NVIDIA Jetson Nano). The equipment also includes required AI software platforms such as TensorFlow, ROS, and MATLAB utilized to train RL models, deploy control, and conduct virtual simulation in Gazebo. Each module is woven into an integrated pipeline to enable the development, test, and verification of the navigation system under different operating conditions. The diagram also illustrates how test environments like traffic simulation, obstacle interaction, and environmental variability are woven with the sensing, decision-making, and actuation modules of the system. This organized simulation infrastructure offers a realistic and scalable environment for testing the efficiency, robustness, and responsiveness of the suggested autonomous robot navigation system.

Figure 7.

The complete simulation and implementation framework integrates hardware components, software tools, and testing environments.

The entire hardware deployment and real-world testing of the outlined robotic system are demonstrated in Fig. 8. The deployment involved building the physical robot body, installing necessary sensors such as LiDAR and cameras, and running the control system on an edge computing environment. RL-based path planning and obstacle avoidance software modules were also deployed and executed in real time. Part of the testing was conducted in the evaluation of the robot within an actual smart campus environment, where it performed well under pressure in navigating through dynamic corridors based on the learned RL policy. The configuration offers evidence of the usability of the framework in the field and robustness in converting between simulated testbed and real-time deployment without model reconfiguration or architecture design alterations.

Figure 8.

Real-time deployment of the suggested robotic navigation framework. The figure indicates key milestones including hardware integration of the robot platform, sensor onboarding (LiDAR and cameras), reinforcement learning model training, and internal edge computing setup.

The robot is deployed on a smart campus corridor environment for real-world navigation performance evaluation. The deployment confirms the capability of the system to transition from simulation to real-world deployment under changing environments.

3.5. Reinforcement learning and path planning

The model leverages the capability of RL in allowing dynamic path planning for real-time decision-making in urban environments. Unlike traditional approaches, this RL-based approach learns adaptability to dynamic conditions, including varying traffic densities, unpredictable obstacles, and adverse weather. The reward function for the model gives precedence to routes that have less travel time, energy consumption, and collision risk; therefore, it provides both safety and efficiency. The RL and path planning behavior of the proposed system is illustrated in Fig. 9. Figure 9(a) shows the evolution of the cumulative reward across episodes, indicating stable convergence of the PPO agent. Figure 9(b) captures intermediate policy shifts, visualizing how the action probabilities evolve as the agent refines its decision-making in response to reward gradients. Figure 9(c) presents a 3D path cost surface, which reflects the learned navigation space where lower valleys represent optimal paths. The composite figure provides an in-depth understanding of how the agent translates real-time sensor inputs into safe, efficient, and adaptive trajectory plans using learned policy distributions.

Figure 9.

Reinforcement learning-based path planning visualization. (a) Reward convergence curve across training episodes. (b) PPO policy distribution shift showing decision refinement. (c) 3D cost landscape revealing optimal trajectory valleys for adaptive navigation.

3.6. Algorithm and flow chart for IoT-Enabled adaptive reinforcement learning navigation

3.6.1. Algorithm for optimal path planning using reinforcement learning

The proposed algorithm introduces a real-time, multimodal RL framework for adaptive robotic path planning that surpasses conventional static or single-objective approaches. The novelty of this work has considered the dynamics of heterogeneous sensorial blend – traffic flow, obstacle positioning, and weather changeableness to form a single state representation – thus the complexity of the actual world itself. Multi-objectives reward model brings a level of depth to path planning since it can enhance the policy of the agent with respect to travel time management, consumption, and collision risk management, as opposed to the other path planning systems that utilize a fixed map or standard cost functional. The following algorithms, as introduced in Algorithm 1, were for autonomous navigation of the robot in a dynamic smart campus and smart urban mobility environment. The algorithm executes the PPO P algorithm on the PPO-data of real-time environmental measurements, T, which includes crowd density (brighter colors) and pedestrian flow (darker, sparsely distributed colors) and obstacle, O, data, trained on with LIDAR, and weather information, W, a sequence of continuous actions (linear velocities +Z and angular velocities Z) estimated by using the following steps: The original Tripartite solution is a graph inspired Earth-weighted graph neural network (GNN) that is trained on the PPO-data of real-time environment measurements, T. Moreover, the algorithm introduces a context-aware reward function that dynamically adjusts weights for travel time, energy consumption, collision risk, and stability risk based on real-time IoT inputs – prioritizing, for instance, collision avoidance in dense crowds or under wet surface conditions. Enhanced by noise filtering, edge-case handling (such as safe stops and dynamic replanning), and path caching for recurring scenarios, this method achieves real-time performance at 10 Hz with a computational complexity of O(nlogn) for GNN processing. Overall, it represents a significant advancement over traditional RL and path planning methods, offering robust, adaptive, and efficient navigation in complex, dynamic environments.

Algorithm 1:

Continuous linear arc transition mode based on the register,

3.6.2 Flowchart of IoT-enabled adaptive RL navigation algorithm

Figure 10 illustrates the workflow of our novel IoT-enabled adaptive RL algorithm for autonomous robotic navigation in smart campus and urban mobility systems. This flowchart delineates the algorithm’s iterative process, integrating real-time environmental data (T) from IoT sensors, obstacle data (O) from LIDAR, and weather conditions (W) to compute an optimal path. The algorithm’s novelty lies in its use of a GNN with dynamic edge-weighting to model spatial–temporal IoT data (e.g., pedestrian flow), a convolutional neural network (CNN) for obstacle mapping, and a context-aware reward function with adaptive weights that prioritize efficiency, energy, collision avoidance, and stability based on IoT inputs (e.g., increasing collision risk weight in dense crowds). Figure 10 captures these innovative components, along with robust exception handling for high-risk scenarios and IoT data loss, providing a clear schematic of the algorithm’s advanced decision-making process for dynamic urban environments.

Figure 10.

Flowchart of IoT-enabled adaptive RL navigation algorithm.

3.6.5. Module interaction overview

The architecture relies on the continuous and dynamic interaction of multiple complementary modules rather than a single intelligence module. The visual streams from cameras and depth maps through the LiDAR are processed in a CNN. This module is fundamentally responsible for identifying road-use, barriers, and lane markings in real time, converting unstructured pixel data into an obstacle map for the purpose of action. At the same time, the spatiotemporal state of the larger context (e.g., pedestrian flows in corridors, traffic congestion at intersections, or changing flows on campus routes) is derived from IoT sensors and represented as a dynamic graph. This representation is run through a GNN, adjusting edge weight based on some level of congestion or crowding, so as to provide the overall system with a macro rather than limited field-of-view representation of the state of the environment. Once generated, the two streams of perception are coupled within the RL controller. Here, the PPO agent evaluates the fused state vector, which combines CNN-based obstacle features, GNN-derived contextual information, and auxiliary weather or terrain inputs filtered through the Kalman fusion layer. The agent does not simply minimize distance or time; instead, it optimizes a multi-objective reward function (as defined earlier in Eq. (1)), which simultaneously balances travel efficiency, power usage, collision avoidance, and stability on uneven or slippery surfaces. Importantly, the weights within this reward adapt as the IoT context changes – for instance, under fog or heavy pedestrian movement, the penalty on collision risk is automatically amplified, while in long-range open navigation, energy efficiency assumes greater priority.

Once the PPO agent proposes high-level waypoints, they are handed over to traditional planners for refinement. Deterministic algorithms such as A* are favored in structured settings, producing near-optimal paths on grid-like layouts, whereas sampling-based methods such as RRT are more effective when the robot is placed in semi-structured or cluttered environments. These classical planners ensure that the abstract policies of the RL agent are translated into executable, collision-free trajectories with low actuation latency. A schematic illustration is included in Fig. 11 to visualize the flow of data and decisions across these modules.

Figure 11.

Integrated interaction of CNN, GNN, PPO, and classical planners.

Finally, these refined trajectories are dispatched to the low-level control system of the robot, where velocity and heading adjustments are executed. Sensor feedback immediately closes the loop: CNNs update obstacle maps, the GNN revises its graph representation, and the PPO agent recalculates policies if environmental states deviate from expectations. In this way, perception, decision-making, and planning are not isolated modules but continuously reinforcing processes.

In essence, CNNs provide fine-scale obstacle intelligence, GNNs supply global situational awareness, PPO adapts control decisions in real time through multi-objective optimization, and A*/RRT guarantee safe and smooth local trajectories. Their integration establishes a hybrid navigation framework capable of handling uncertainty, dynamic obstacles, and context variability while remaining deployable on embedded edge hardware.