Self-Navigation Systems for Drones

A hierarchical architecture for controlling autonomous vehicles that improves performance and robustness by breaking down the vehicle control stack into multiple layers. The layers have separate planning and checking functions to handle tasks at different time scales and complexities. The layers receive inputs from above and send outputs to below. Each layer can override the one above if the planning conflicts. The lower hardware layer converts control signals. The upper layers plan tasks like trajectory generation. The layers in between check and intervene based on scene data. This allows complex tasks to be planned slower than simple ones.

Source

Automated generation of safer drone delivery routes to enhance safety of the drones and the parcels they carry. The method involves using machine learning to analyze safety factors like bird density, road density, water bodies, etc. for different flight paths. A matrix representing safety scores for sub-regions is generated. During flight planning, an AI model evaluates routes based on this matrix to recommend safer alternatives.

Source

Approximate algorithm for finding shortest paths and closest points in 3D obstacle environments with applications to drone navigation. The algorithm uses an epsilon graph data structure to efficiently find k closest points to a query location in the presence of obstacles. The epsilon graph is created by grouping arrival points, obstacles, and valid nodes/edges based on proximity. It includes anchor nodes for each arrival point that are reachable through valid edges. This allows finding k closest points by starting from the query point and traversing the epsilon graph.

Source

Localization technique for autonomous vehicles that improves accuracy and robustness by fusing pose estimates from multiple sensors like IMU, wheel encoders, LIDAR alignment, and lane alignment. It uses a Kalman filter to integrate the sensor data into a single pose state representing both local and global position. This allows out-of-order measurement processing and safeguards against occasional sensor error.

Source

Validating localization systems of autonomous vehicles by simulating rare and unusual environmental conditions not commonly encountered in the real world. The method involves augmenting real-world log data with simulated anomalies like sensor failures, degraded map data, occlusions, and latency. The augmented data is then used to simulate the localization system's operation and determine metrics like accuracy and robustness. This allows evaluating and improving the localization system's performance in extreme conditions that may never be seen in normal operation.

Source

Safe reinforcement learning algorithm called shield-DDPG for UAV path planning in urban airspace that can avoid collisions and ground impacts while ensuring fast response to UAV missions. The algorithm uses a shield module to prevent unsafe actions and a safety specification to ensure safe behavior. It converges quickly by enforcing safety constraints throughout training. The shield module is based on linear temporal logic and checks if actions maintain safe distances from obstacles and the ground. This prevents unsafe actions from being learned by the RL algorithm. The shield-DDPG algorithm provides a reliable and safe UAV path planning method for urban airspace.

Source

Using path signatures to enable autonomous vehicles to track and follow a reference path. The technique involves calculating a path signature for the vehicle's historical path, observing the current state, updating the signature, and identifying actions to follow the reference path based on the updated signature. This allows efficient path planning and trajectory generation compared to full re-planning from scratch.

Source

Real-time path finding for autonomous vehicles that improves path calculation efficiency and memory usage by leveraging deep learning to intelligently sample and prune the search space. The method involves embedding past path and sensor data into a topology-based neural network, which outputs likely path areas at the current time. This allows focusing sampling and pruning efforts on areas with higher path probability. By dynamically updating the topology based on new obstacles, it reduces unnecessary node expansion and memory use. The network is trained to accurately predict the path distribution based on cross-entropy loss for sampling areas and mean squared error loss for path prediction.

Source

A system for managing unmanned vehicles like drones at vertiports, such as hotels, buildings, etc. The system includes charging and repair stations, automated loading/unloading, autonomous navigation, and real-time tracking. It also leverages AI/ML, blockchain, and smart contracts for optimized operations, data sharing, and security. The vertiport system enables efficient and scalable urban aerial logistics using autonomous drones.

Source

An inertial measurement unit (IMU) that adaptively filters its sensor output to improve accuracy in dynamic conditions. The IMU has a rule-based filter selection mechanism that determines the optimal filter parameters for each sensor based on the current application and environment. The rules may define thresholds or conditions for filter selection. This allows the IMU to dynamically adapt the filtering to provide output sensor data that is optimized for the specific use case and conditions.

Source

Improving unmanned aerial vehicle (UAV) functionality like navigation, object recognition, and obstacle avoidance by combining data from different types of sensors like vision and proximity. The UAV carries multiple sensors like cameras and ultrasonic sensors. It uses them to generate an environmental map by fusing the data. This map provides accurate location info and obstacle positions. The UAV can then use the map for tasks like autonomous flight, return to base, and obstacle avoidance. The sensor fusion improves accuracy compared to using just one sensor type.

Source

Visual target recognition and positioning method for quadrotor drones using two pairs of binocular cameras to improve accuracy, coverage, and reliability in complex environments. The drone has one pair of binocular cameras at the front and another pair at the rear. This provides 360-degree coverage for comprehensive spatial perception. Real-time target recognition and positioning is achieved using high-performance computing and GPU acceleration. The binocular cameras improve accuracy by resolving blind spots, handling lighting variations, and reducing structured light interference.

Source

Enabling autonomous aerial navigation for drones in low light conditions without disabling obstacle avoidance. The drone uses a learning model trained on simulated infrared images for obstacle avoidance in night mode. In day mode, the drone still uses the cameras but filters out the infrared data to improve image quality for navigation. This allows the drone to autonomously navigate in low light without relying solely on GPS. It prevents disabling obstacle avoidance or manual navigation in low light environments.

Source

Visual navigation of unmanned aerial vehicles (UAVs) using liquid neural networks to enable more accurate and responsive autonomous flight in complex environments. The method involves capturing images from the UAV's camera, feeding them through a liquid neural network to process and analyze the environmental information, and using the network's output to generate new desired flight controls. This allows the UAV to navigate through dynamic and unknown conditions more effectively compared to traditional control methods. The liquid neural network is implemented using a software control system with threads for image processing and reasoning. The network runs on a GPU using Tensorflow for accelerated inference.

Source

Real-time, accurate flight path navigation for drones in complex environments using fused sensor data. The method involves simultaneously obtaining positioning data from GPS and image data from onboard cameras. Preprocessing is done on the data in an edge computing environment using optimization algorithms to reduce complexity and improve speed. Fusion of the preprocessed data is then done with dynamically adjusted weights based on environmental conditions. The fused data is used for obstacle detection and drone navigation.

Source

Autonomous navigation system for small quadcopter drones in dynamic environments using onboard sensors and processing. The system allows the drone to safely navigate complex environments with moving obstacles without external sensors. It uses a depth camera to create an occupancy grid map of the environment, extracts independent moving obstacles, estimates their motion, and applies an improved collision avoidance algorithm called ORCA to plan a smooth, collision-free trajectory to the target point. The ORCA algorithm generates optimal mutual collision avoidance sets for multiple moving obstacles.

Source

Obstacle avoidance method for drones that improves safety and autonomy by using machine learning and a hierarchical response strategy. The method involves monitoring flight status, predicting future trajectories using machine learning, and responding appropriately to anomalies based on severity. This allows intelligent obstacle avoidance, optimized flight efficiency, and quick response to emergencies while reducing overreaction to minor anomalies.

Source

Indoor drone target tracking system using depth cameras and SLAM for following targets in indoor environments where GPS is not available. The system leverages onboard cameras and IMU to track targets without relying on external positioning systems. It integrates target detection, tracking, SLAM, and path planning to enable real-time indoor drone following without GPS.

Source

Vision-based autonomous landing method for fixed-wing drones near the ground when GPS and inertial navigation fail. The method involves real-time identification and tracking of runway lines and horizons in images, then calculating the drone's pose using photographic geometry to land accurately. This allows autonomous landing with just camera inputs when traditional navigation fails.

Source

Detecting obstacles ahead of a drone using high-precision maps during flight to allow early obstacle avoidance preparation. The method involves periodically checking within a distance called the status adjustment distance if any obstacles exist based on high-precision map data. If obstacles are found, the drone transitions to obstacle avoidance flight mode. This allows the drone to prepare for obstacle avoidance regardless of user settings since it can detect obstacles earlier using maps instead of radar.

Source