Reinforcement Learning

Project Goals

Our team used reinforcement learning frameworks and techniques to train a quadcopter to fly in a simulated indoor environment. This type of machine learning approach enables the agent to learn by trial and error and act under various surrounding complexities. We used Unreal Engine and Microsoft’s AirSim, which acts as the interface between our python code and the simulator.

Solution Design

The drone had multiple front cameras that helped our agent make decisions based on the image depth (how far the drone/agent is from an object).  This information was then fed into an algorithm called Q-learning, which provides positive reward if the drone does not collide, and a negative reward if it gets really close or clashes with an object/wall.


After 100 hours of training, our agent was able to navigate itself in an indoor environment without colliding for a few minutes.

Next Steps

  • Make the drone reach a specific target in a space

  • Build a more dynamic reward system that will help the agent navigate the environment and reach the target more effectively

  • Run the solution into multiple environments

  • Build an environment where multiple agents can work together to reach a target

We deliver actionable insights by empowering your organization with AI so you can focus on business growth.

We offer the leading edge end-to-end AI solutions, enabling every employee, customer, and citizen with sophisticated AI technology and easy-to-use AI applications. With our solutions, our partners can become leaders in their respective domains by becoming more innovative in serving their customers, collaborating among all key stakeholders in extracting business value and reducing costs.