Perception & SLAM
SLAM (Simultaneous Localization and Mapping)
SLAM is a technology used in robotics and autonomous systems to simultaneously keep track of the robot’s pose and build a map of an unknown environment, especially in environments where GNSS cannot be used. To achieve this, it integrates measurements from diverse sensors such as cameras, LiDAR, Radar, and IMUs.
VIO (Visual-Inertial Odometry)
Cameras are lightweight and compact sensors that capture rich visual information, providing dense texture and feature representations of the environment. Unlike active sensors, they excel at recognizing distinct visual landmarks necessary for loop closure and relocalization. VIO combines this visual stream with high-frequency inertial data to track the robot’s pose. This fusion compensates for motion blur during rapid movements and enables precise navigation in GPS-denied areas, such as indoor spaces or urban canyons.
LIO (LiDAR-Inertial Odometry)
LiDARs are active sensors that generate precise 3D point clouds, offering accurate geometric information about the surroundings regardless of lighting conditions. They provide a wide field of view and long-range detection, making them ideal for mapping large-scale environments. LIO tightly couples these dense geometric measurements with inertial data to estimate trajectory. This combination ensures high-accuracy localization and mapping, establishing it as a standard solution for autonomous driving and mobile robotics.
RIO (Radar-Inertial Odometry)
Radars are powerful sensors that can measure distances and estimate radial velocities with high precision, even in challenging environmental conditions where cameras and LiDARs might fail. They are unaffected by changes in illumination, robust against repetitive structures, and can operate effectively in environments with smoke or fog due to their long wavelengths. RIO integrates radar data with inertial data from IMUs to estimate the position and orientation of a robot, ensuring reliable operation even when optical sensors are compromised.
Legged Odometry (Kinematic-Inertial)
Legged robots possess unique proprioceptive capabilities, utilizing joint encoders and contact sensors to sense their physical interaction with the terrain. This method is independent of external environmental features like texture or lighting, making it fundamentally different from vision or laser-based approaches. Legged odometry integrates leg kinematics and contact states with inertial data to compute the robot’s motion. This provides a robust fallback layer for state estimation, maintaining tracking even in featureless environments or when external sensors are obstructed.
Gaussian Process Spatial Modeling
Gaussian Process (GP) regression offers a robust framework for spatial mapping tasks, including height mapping and gas source estimation. By modeling spatial correlations, it enables the smooth reconstruction of continuous surfaces from sparse and noisy sensor measurements. Uniquely, this method not only predicts values but also quantifies uncertainty, providing critical information for risk-aware robotic decision-making. Consequently, it ensures reliable environmental modeling in complex scenarios involving uneven terrain or diffusive gas distributions.
Robotic Perception for Precision Agriculture
Multi-modal perception technology is essential for robotic agriculture systems to accurately monitor crops in confined indoor environments. By integrating high-resolution cameras with LiDAR sensors, this approach overcomes the limitations of conventional sensing to achieve sub-centimeter precision. This high-fidelity data enables automated tasks such as individual fruit counting, yield estimation, and detailed growth analysis. Consequently, it provides the reliability required for manipulators and drones to operate autonomously in modern smart farming applications.
Neural network compression
Neural network compression plays a pivotal role in enhancing the utility of mobile robotics by enabling efficient deployment of deep learning models on resource-constrained devices. In the realm of mobile robotics, where computational resources and power consumption are critical constraints, compressed neural networks facilitate real-time decision-making, navigation, and perception tasks. By reducing the size and complexity of neural networks through techniques like pruning, quantization, and knowledge distillation, mobile robots can efficiently process sensory data, navigate dynamic environments, and execute tasks with improved speed and accuracy. Neural network compression empowers mobile robotics applications by enabling them to operate seamlessly in real-world scenarios while conserving computational resources and energy.
Formation-based Human Detection System
Integrated UAV formation flight systems leverage visible-thermal fusion to overcome the limitations of single-sensor detection in complex search and rescue (SAR) environments. While RGB cameras provide high-resolution visual details, thermal IR sensors are essential for identifying human heat signatures in low-visibility conditions or dense foliage where visual cues are obscured. By combining this robust sensing with real-time obstacle avoidance and coordinated communication, these systems ensure reliable human detection and autonomous navigation in challenging terrains like forests and mountains.