AI-Driven Visual Navigation for Smart Lab Tour Guide Robot

While robotics and automation lab technologies provide a learning platform, the frequent human involvement in guided lab tours highlights technological inefficiencies [1]. Traditionally, line-following robots have been used to guide people. However, these types of robots are resilient to environmental changes. Line-following robots use predefined paths and are not flexible to adapt to environmental changes [2]. These contradictions are a major obstacle to their use in real laboratories. It is difficult to use line-following robots in places where equipment positions change and tourist traffic patterns change frequently [3]. To overcome these challenges, this research uses an AI-driven visual navigation system. The aim of this research is to develop a navigation system for laboratory environments without predefined paths [4]. The architecture is developed by three technologies: visual simultaneous localization and mapping (SLAM) for spatial awareness, deep learning-based object detection for obstacle recognition, and adaptive path planning for intelligent navigation decisions.

ORB-SLAM2 has real-time localization and mapping in its monocular configuration [5]. It provides metric scale environmental perception. In this research, dynamic obstacle detection is performed by YOLOv5s [6]. This YOLOv5 architecture is specifically fine-tuned for laboratory environments. An A* algorithm integrates real-time obstacle information with global path planning in the navigation system [7]. It enables dynamic rerouting in response to environmental changes [8]. This research contributes to a novel sensor fusion framework that combines visual SLAM features with deep learning-based object detection. It creates a unified representation for navigation decision making. The hardware is Raspberry Pi 4 for optimized implementation of computationally intensive algorithms. The Raspberry Pi 4 has achieved real-time implementation through careful system design and optimization. This research provides experimental validation in realistic laboratory settings. Several metrics include navigation performance to characterize system performance. Computational requirements and environmental robustness should be considered in developing the framework.

Section 2 reviews relevant literature on visual SLAM, deep learning for robotics, and path planning algorithms. Section 3 details our system architecture and mathematical framework. Section 4 describes the implementation methodology. Section 5 presents experimental results and analysis. Section 6 concludes with future research directions.

3.1 System overview

In this research, SLAM and obstacle detection are performed using a camera as the primary sensor. Raspberry Pi is used as the main processing unit in this embedded system. In the research, localization and mapping are performed with the help of ORB-SLAM2. Planning and object detection are performed using YOLOv5. Raspberry Pi handles the high-level navigation commands and is implemented by Arduino along with motor control. It has a hierarchical control structure in which high-level decision making is implemented by the low level activation system.

3.2 Robot kinematics

Different configurations are commonly used in mobile robotics. This given kinematic model provides a robust underlying mathematical model for motion control and path following. The robot’s pose in the world coordinate system is represented by $q=[x, y, \theta]^T$, where x and y denote position coordinates and $\theta $ represents orientation. The kinematic equations are:

$\dot{q}=\left[\begin{array}{c}\dot{x} \\ \dot{y} \\ \dot{\theta}\end{array}\right]=\left[\begin{array}{c}v \cos \theta \\ v \sin \theta \\ \omega\end{array}\right]$            (1)

For a differential drive robot with wheel radius r and wheel separation L, the linear velocity v and angular velocity ω relate to wheel angular velocities ${{\omega }_{r}}$ and ${{\omega }_{l}}$ as:

$v=\frac{r}{2}\left( {{\omega }_{r}}\ +\ {{\omega }_{l}} \right),\ \omega =\frac{r}{L}\left( {{\omega }_{r}}-\ {{\omega }_{l}} \right)$           (2)

These equations form the basis for our motor control system implemented on the Arduino microcontroller, converting high-level velocity commands into precise wheel movements.

3.3 Visual SLAM implementation

We implement ORB-SLAM2 in its monocular configuration, optimized for embedded systems. The SLAM problem is formulated as maximum a posteriori (MAP) estimation:

$x_{o,t}^{*},~{{m}^{*}}~=~\text{arg}\underset{xo:t,~m}{\mathop{\max }}\,p({{x}_{0:t,}}m|~{{z}_{1:t}},~{{u}_{1:t}})$           (3)

Applying Bayes’ rule and Markov assumption:

$\begin{aligned} & p\left(x_{0: t}, m \mid z_{1: t}, u_{1: t}\right) \propto \\ & p\left(x_o\right) p(m) \prod_{k=1}^t p\left(z_k \mid x_k, m\right) p\left(x_k \mid x_{k-1}, u_k\right)\end{aligned}$            (4)

Known parameters from laboratory equipment are taken for monocular SLAM. The G2O framework and ORB-SLAM2 are combined for loop closure detection and pose graph optimization. This is useful for correcting accumulated drift over time and maintaining global consistency in the system. The system runs on a Raspberry Pi 4 at 30 Hz and tracks approximately 1000 ORB features per frame.

3.4 YOLOv5 object detection integration

In this research, YOLOv5s was used as an optimizer for embedded deployment. To find the tuned model, more than 500 images of a laboratory with various obstacles such as chairs, equipment cards, and some human figures were taken. The bounding box and confidence score were found to find the output for each frame:

$Confidence=P(Object)\times IOU_{pred}^{truth}$             (5)

Detected obstacles are converted into a local occupancy grid with resolution 0.1 m using the following transformation: for each detection with bounding box (${{x}_{c}},{{y}_{c}},w,h$) and confidence c, we compute the corresponding occupancy probability ${{p}_{occ}}=c.~{{\text{e}}^{-\frac{{{d}^{2}}}{2{{\sigma }^{2}}}}}$, where d is the distance from grid cell center to detection center and σ = 0.3 m. The detection system operates at 15 Hz on the Raspberry Pi 4.

3.5 Enhanced A* path planning

The path planner operates on a 2D grid map generated by SLAM with resolution 0.05 m. The A* algorithm minimizes:

$f(n)=g(n)+h(n)$            (6)

where, g(n) is the actual cost from start to node n, and h(n) is the heuristic estimate to goal. We employ Euclidean distance as heuristic:

$h\left( n \right)=\ \sqrt{{{({{x}_{n}}-{{x}_{g}})}^{2}}+{{({{y}_{n}}-{{y}_{g}})}^{2}}}$         (7)

A real-time cost map is used to enhance the algorithm to avoid moving obstacles. A safety radius of 0.5 meters is maintained around obstacles. It is determined as optimal through experimental validation to balance safety and navigation efficiency. The algorithm attempts to plan a new route at 5 Hz when new obstacles are detected within a range of 2 meters.

5.1 Navigation performance analysis

Table 1 gives a comparison of AI based SLAM system and traditional line following system. This AI based SLAM system is better in all parameters except computational load. It reduces the time by 40% in two times and it shows that this system is more efficient than traditional system. The developed system reduces the deviation by 65.6% i.e. from 3.2 cm to 1.1 cm. It shows that SLAM provides excellent localization accuracy. This SLAM tracking provides localization success rate of 98.2%. In this system, primary failures occur at 15% turns and 72% in low texture areas. Illumination changes are causing 13% failure results. Regression analysis is used to find the correlation between path deviation and SLAM tracking stability. A strong negative correlation R² = 0.83 was found between SLAM feature number and path segmentation. If ORB feature tracking is reduced to less than 300 features per frame, the path deviation increases by an average of 2.7 cm. This shows that sufficient feature density is important for navigation accuracy.

Table 1. Navigation performance comparison (100 trials)

5.2 Obstacle avoidance performance

Table 2 shows obstacle awareness performance is demonstrated after 200 trials. The system maintains a high success rate in challenging situations. The computational load is increasing due to the crowded environment, thus reducing the system performance. This is due to the more frequent path lab planning requirements. A success rate of 91.8% has been achieved in this situation. The ultrasonic sensor is providing emergency stop functionality. It provides 100% liability for immediate collision prevention. Parameter sensitivity analysis is performed to understand the relationship between YOLO confidence threshold and avoidance performance.

Figure 5 shows three confidence shots reduced from 0.7 to 0.5. It shows an 18% increase in detection rate but at the same time a 7% decrease in avoidance success rate. This is due to increased false positives. The analysis gives an optimal threshold of 0.65. I will discuss the balance accuracy of 92.3% and recall of 89.7%. The abstract avoidance success rate strongly suggests that R² is equal to 0.76 with the time available for replanning. When comparing obstacles encountered within a 1.5 m range of the robot to obstacles encountered beyond 2.5 m, the avoidance success rate is 15% lower.

tu_pian_16.png

The optimal threshold of 0.65 balances detection rate, precision, and avoidance success

Figure 5. Parameter sensitivity analysis showing trade-offs in YOLO confidence threshold selection

Table 2. Obstacle avoidance performance (200 trials)

5.3 Computational performance and system reliability

Figure 6 presents a comparison of computational resource utilization between AI SLAM and LF systems. Four parameter values namely CPU usage, memory usage, inference time and power consumption are reported. AISLAM inference time includes both SLAM and YOLO processing time. The CPU load utilization shows 78% utilization during navigation. It starts from 92% during simultaneous SLAM mapping and obstacle detection. Some important consequences of this high utilization are given below.

1. Thermal Management: The Raspberry Pi 4 reached 85°C after 25 minutes of operation without active cooling. Due to this, it reduces the CPU frequency from 1.8 GHz to 1.4 GHz. This frequency reduction increases the SLAM processing time. It increases the processing time by 35% and, therefore, causes frame drops in 2.1%.

2. Real-time Performance Impact: Simultaneous loop closure and solid obstacle detection in SLAM occur during peak load periods. The total processing latency for 15Hz operation is sometimes more than 66ms per frame. SLAM tracking loss recovers to 0.8% of frames in 100ms processing. YOLO frame dropping is 1.3% in dense obstacle conditions. This causes a 120ms delay between path planning re-planning events.

3. Reliability Concerns: Memory fragmentation increases SLAM initialization time by 45% in long-term testing over 10 hours. Periodic memory cleanup routines are required to maintain system stability.

4. Frame Drop Analysis: When SLAM and YOLO are under heavy load conditions, it shows 67% frame drop in the analysis. ORB-SLAM 2 has given 33 events during loop closure optimization. It sometimes monopolizes CPU resources for 150 to 200ms.

tu_pian_17.png

AISLAM inference time includes both SLAM and YOLO processing

Figure 6. Computational resource utilization comparison between systems

5.4 Environmental adaptability

The system works as shown in Table 3 of different environmental conditions available. The degradation of performers in low light conditions highlights the limitations of visual perception. This maintains good performance in high-traffic conditions with a success rate of 91.8%. This shows how effectively the system handles complex environments. It detects 70% localization failures in areas with low texture. This happens when reduces the ORB feature count to less than 200 features per frame. In low light conditions, there is noise in the camera which reduces the feature mapping accuracy by 40%. Five pixel tracking loss in ORB SLAM2 is due to faster motion blur. The key relationships have been identified as follows:

1. SLAM features vs. localization error: Exponential decrease in error with increasing feature count (error e−0.002features)
2. Detection distance vs. avoidance success: Linear decrease in success rate as detection distance decreased (success rate = 0.95 - 0.08 × distance, for distance in meters)
3. CPU load vs. response time: Quadratic increase in planning latency with CPU utilization (latency 0.02 × CPU²)

Table 3. Performance under varying environmental conditions with failure mode analysis

5.5 User experience evaluation

Table 4 considers 50 of the participants for the user evaluation results. It is found that users consistently rated the AI-SLAM system for smooth navigation and obstacle avoidance. The time ratings obtained show that the increased computational complexity does not negatively affect the participants. The correlation between usability satisfaction and navigation smoothness is R = 0.78.

This suggests that reducing the startle moment during obstacle avoidance. Users are sometimes tolerant of poses due to sudden changes in direction. This indicates that the high response time ratings sometimes result in computational delays.

Table 4. User experience evaluation (N = 50)

5.6 Technical contributions

This research demonstrates several important technical benefits. This research successfully implements the integration of embedded hardware with real-time visual SLAM, deep learning and path planning. A novel sensor fusion framework is integrated with YOLO detection in dynamic devices along with ORB features. The implementation of this model provides adaptive navigation without predefined paths. It is providing a fundamental change from the traditional approach to system modification. This capability of the system provides deployment of the system in various environmental conditions without expensive infrastructure. This research has achieved a balance between computational requirements and practical performance. The cost-effective hardware demonstrates that it can be easily implemented in the real world.

5.7 Limitations and challenges

This research found some limitations. The first important limitation is that excessive lighting provides low performance. The CPU utilization is 78% indicating that the computational intensity can limit the deployment. This indicates that the system requires a more resource-constrained platform. Scale in the monocular SLAM presents challenge for metric navigation. The limited ability to predict obstacle motion patterns offers another area for improvement. This research found that purely reactive obstacle avoidance is insufficient in highly dynamic environments. Recent approaches like DynaSLAM [13] and DS-SLAM [14] have shown promise in handling dynamic scenes through semantic segmentation and inpainting techniques. However, these methods are computationally intensive and may not be suitable for embedded systems like Raspberry Pi 4 without significant optimization.

5.8 Comparison with alternative approaches

Similar approaches are available for many cameras, including cost-effective solutions suitable for academic and laboratory applications. The performance between computational cost and the navigation capabilities of the system provides a strategic balance for real-world applications. In most situations, absolute performance is less important than reliability and affordability.

If we compare this system to more complex systems like DyGS-SLAM [15] which offers double-constrained optimization for better map reconstruction in dynamic scenes, or deep reinforcement learning approaches [16] that provide more adaptive navigation policies, this approach shows that autonomous navigation capabilities are achieved with simple hardware for advanced robotics technologies. The trade-off is between computational complexity and practical deployment feasibility.