Motion Planning for an UAV inside narrow corridors


In this project, we design motion planning algorithms for an UAV flying indoors. The planner has to navigate the UAV through tight corridors as well as be robust against localization errors in this GPS denied environment.


In GPS-denied environments with large state estimation uncertainty and small obstacle free space, producing a reliable collision-free trajectory is very difficult. The major issues in such scenarios are limited sensor range, large drifts in state and presence of irregular clutter. A suitable trajectory planner must be able to compute useful trajectories at a high frequency. Conventional optimal planners are not suitable because the limited sensor range (3.5 metre) is not a large enough space to have a meaningful optimal solution. Drift in state implies that planning horizon has to be very limited.
A reactive planning approach, where the robot reacts to immediate obstacles, is very effective in keeping the robot safe at all times and avoiding obstacles of arbitrary complexity. However, getting such methods to plan in between clutter and progress towards a goal is often tedious.
We present a solution which balances reactivity and optimality, thus effectively avoiding clutter while progressing towards goal. The system relies on predefined behaviours such as corridor following and turning, thus being capable of tolerating large drifts in state. The predefined trajectories are passed through a fast optimizer, enabling the robot to be reactive and safe.


The schematic of the planning system is shown below

An example of an initial guess and fast optimization is shown below


Inorder to gain information about the environment, for aiding localization and obstacle detection, the view of the sensor must be optimized as shown in the schema

A demonstration of the plan for an UAV moving inside the university. The UAV has a kinect on board, which is used for state estimation and mapping.

Example 1. Localization fails at the last turn due to drift in state, but planner continues to work. Robot reduces uncertainity at the junction helping robot to localize correctly globally.

Example 2. Localization fails at second turn, but planner keeps robot safe. On recovery of global localizer, robot resumes correct path.

Example 3. In this example state is degraded artifically by inducing drift, however planner keeps the robot safe and completes mission.