YOLO Object Detection

R-CNN is a two-stage detector that is generally more accurate but significantly slower. YOLO is a single-stage detector that prioritizes inference speed. For mobile robots that need to react to obstacles in milliseconds, YOLO is preferred because it can run at high frame rates (30-60 FPS) on embedded hardware, whereas R-CNN might struggle to reach real-time performance.

While lightweight versions (like YOLO-Nano) can run on CPUs (like Raspberry Pi), industrial AGVs typically use Edge AI accelerators. The NVIDIA Jetson series (Nano, Orin, Xavier) are the industry standard, providing the CUDA cores necessary to run full YOLO models with low latency and reasonable power consumption.

YOLO relies on visual data, so performance degrades in darkness. However, this is mitigated by training the model on low-light datasets or using image augmentation techniques (like adjusting gamma/brightness) during training. For total darkness, YOLO can be combined with thermal cameras or active IR illumination.

Yes, this is a core strength of the architecture via "Transfer Learning." You can take a pre-trained YOLO model and fine-tune it with a smaller dataset of your specific pallets or products. This process is much faster than training from scratch and yields high accuracy for custom facility objects.

YOLO provides a 2D bounding box on the image, but not depth directly. Distance is usually calculated by fusing YOLO output with data from a LiDAR sensor or a stereo depth camera (RGB-D). Alternatively, if the object size is known, the bounding box size can be used to mathematically estimate distance.

Yes, there are widely supported packages (like `darknet_ros` or `ultralytics_ros`) that wrap YOLO inference engines into ROS nodes. These nodes publish detection topics containing class IDs and bounding box coordinates, which the robot's navigation stack can immediately subscribe to for path planning.

Ghost detections can cause robots to stop unnecessarily. To mitigate this, we use "confidence thresholds" (ignoring predictions below 50-60%) and temporal consistency checks (an object must be detected in several consecutive frames to be considered real). Sensor fusion with LiDAR also helps verify if a visual detection actually has physical mass.

For most industrial robotics applications today, YOLOv8 is a strong contender due to its balance of speed, accuracy, and ease of deployment via the Ultralytics framework. However, lighter versions like v5-Nano are still excellent for extremely constrained hardware where battery life is the primary concern.

Running deep learning models is computationally intensive. An NVIDIA Jetson Orin running YOLO might draw 15W-40W depending on the mode. While this is significant, it is usually a small fraction compared to the robot's drive motors. The safety and efficiency gains usually outweigh the slight reduction in runtime.

YOLO historically struggled with small objects, but recent versions utilize feature pyramid networks to improve this. Detection depends heavily on camera resolution and distance. For reliable small-object detection (FOD), the camera must be angled downwards or the robot must move slower to allow for higher resolution processing.

No. Inference is performed entirely "at the edge" (on the robot's internal computer). Internet is only required if you wish to upload gathered images to the cloud for retraining the model later to improve performance over time (a process called Active Learning).

The software (YOLO itself) is open-source. The costs come from hardware ($300-$1000 for the compute module + cameras) and the engineering time to label data and integrate the system. Compared to 3D LiDAR safety systems (which can cost $2k+), vision-based navigation is often more cost-effective.