Accurate phase estimation in legged robots, determining whether legs are swinging or in ground contact, is crucial for stability and performance, especially on uneven terrain and with dynamic obstacles. This paper presents a novel approach to phase estimation using signal imaging and lightweight deep neural networks. Time-series data from an inertial measurement unit (IMU) and joint encoders are fused into phase images in real-time using Stacked Channels Image (SCI) and Recurrence Plot (RP) encoding methods. These phase images are then processed for phase classification, eliminating the need for force measurement sensors, which are prone to noise and failure. Extensive experiments on a 3-DoF hopping robot leg, with contact sensor data as ground truth, show that EfficientNet-B0 deep model achieves 98.97% accuracy with RP imaging and an optimized inference time of 2 milliseconds. Comparative validation on an independent dataset from the MIT Mini Cheetah traversing eight terrains demonstrates our model's generalizability. The EfficientNet-RP-Mode1 configuration achieves 96.51% accuracy, outperforming traditional model-based methods and matching or surpassing other deep learning-based approaches, despite requiring fewer sensor inputs. This research highlights the efficiency, robustness, and adaptability of the proposed method, offering a reliable solution for phase estimation in diverse robotic platforms and environments.
Snow accumulation on photovoltaic (PV) panels can cause significant energy losses in cold climates. While drone-based monitoring offers a scalable solution, real-world challenges like varying illumination can hinder accurate snow detection. We previously developed a YOLO-based drone system for snow coverage detection using a Fixed Thresholding segmentation method to discriminate snow from the solar panel; however, it struggled in challenging lighting conditions. This work addresses those limitations by presenting a reliable drone-based system to accurately estimate the Snow Coverage Percentage (SCP) over PV panels. The system combines a lightweight YOLOv11n-seg deep learning model for panel detection with an adaptive image processing algorithm for snow segmentation. We benchmarked several segmentation models, including MASK R-CNN and the state-of-the-art SAM2 segmentation model. YOLOv11n-seg was selected for its optimal balance of speed and accuracy, achieving 0.99 precision and 0.80 recall. To overcome the unreliability of static thresholding under changing lighting, various dynamic methods were evaluated. Otsu’s algorithm proved most effective, reducing the absolute error of the mean in SCP estimation to just 1.1%, a significant improvement over the 13.78% error from the previous fixed-thresholding approach. The integrated system was successfully validated for real-time performance on live drone video streams, demonstrating a highly accurate and scalable solution for autonomous snow monitoring on PV systems.
25 March 2026 in IEEE Robotics & Automation Magazine