UAV-Guided UGV Navigation

UAV-aided mapping and localization enabling a UGV to autonomously traverse mountainous terrain — our solution to the DRDO Challenge at Inter IIT Tech Meet 10.0.

This is our solution to the DRDO Navigation Challenge at Inter IIT Tech Meet 10.0. In the problem statement, we were asked to use a UAV to explore and map a mountainous terrain during summer. After mapping, the UAV must guide a UGV to traverse the terrain during winter, when the roads are covered with snow.

Key points from the problem statement:

  • The UAV has an IMU, a GPS, and an RGB-D camera as sensors.
  • The UGV has no sensors.
  • The mapping must be completed in the textured world, while the UGV navigates in the untextured world.

Code · Slides · Report

Approach

The project is divided into three parts:

  • Mapping — we segmented roads from the UAV camera feed using U-Net and used frontier exploration to map the entire terrain.
  • UAV localization and control — we used ArduPilot coupled with our own control and planning algorithms; localization fused GPS and IMU data.
  • UGV localization and control — we used YOLOv5 on the UAV camera feed to localize the UGV relative to the UAV, and the se2_navigation and navfn ROS packages for path planning and control.

UAV Localization

We fused the incoming GPS and IMU data using an Extended Kalman Filter to produce odometry. The ekf_localization node in the robot_localization package provided 15-DoF odometry. The maximum observed error after multiple goal points was ±0.5 m in all three axes.

Terrain Mapping

Road Segmentation

We prepared a dataset of simulated environment images, manually annotated them using CVAT, and trained a U-Net to segment roads. With standard augmentation, we obtained ~96% accuracy on test data.

U-Net segmentation results (left) and the exploration pipeline (right).

Mapping was implemented using the RTAB-Map package, which uses an RGB-D SLAM approach to create a 2D occupancy grid from the 3D point cloud captured by the RGB-D camera atop the drone. We then implemented frontier exploration using the frontier_exploration package to explore the terrain — moving to new frontiers (boundaries between open and unexplored space) until none remain.

UGV Navigation

Detection and Tracking

Detection and tracking pipeline.

Controls

For UGV control we used a Pure Pursuit controller — a path-tracking algorithm that computes the angular velocity command to move the robot from its current position toward a look-ahead point in front of it.

Planning

  • Local planner — OMPL (Open Motion Planning Library), a collection of state-of-the-art sampling-based motion-planning algorithms.
  • Global planner — TEB (Timed Elastic Band), which locally optimizes the robot’s trajectory with respect to execution time, separation from obstacles, and compliance with kinodynamic constraints at runtime.

We used the se2_navigation package, which subscribes to the car’s odometry (w.r.t. the drone’s initial pose). When a goal point is published in the same odom frame, the SE(2) OMPL planner plans a trajectory — a sequential array of poses — for the car to follow.