The future of drones in construction

As it becomes more evident to general contractors that autonomous drones can efficiently survey land and inspect infrastructure without pilots on the ground, the global drone construction market is expected to grow into a nearly $20 billion industry. Kenji Shimada aims to make this ever-growing industry safer by eliminating concern about onsite drone collisions.

Navigating dynamic environments

In order to avoid a collision, drones must be two steps ahead of the objects around them. This starts by ensuring the drone can “see” everything in its path. Shimada, a professor of mechanical engineering, does this by installing both a high-quality camera and radar sensor onto the drones. This allows the drone to “see” 3D objects around it and understand its distance from each object and person. Because the drone is collecting data from different sensors, each data set can be cross-checked for accuracy, which adds another layer of security.

Next, Shimada ensures that his drones can predict the course of the people around it using a model called the Markov Decision Process. Instead of predicting a single trajectory per obstacle, Shimada’s module generates all possible trajectories including stopping, turning, and forward movement. After processing this information, Shimada’s drones can plan the best route to avoid collision.

As an added bonus to general contractors, Shimada found that his radar sensor is able to measure the coordinates of a construction site after just one thirty second flight.

“Land surveying requires expensive equipment and leaves room for human error,” said Shimada. “Automating this process would save time and money.”

From simulation to real-world application

Reinforcement learning is a machine learning technique that mimics the trial-and-error method people use as they learn to navigate the world. In the real world this would look like a handful of drones flying into people and buildings as they learn how to avoid them. Thanks to digital twins, Shimada’s team is able to expose thousands of drones to any number of collision risks in a destruction-free, injury-free environment.

Oftentimes when a robot trained in simulations is put to the test in the world it fumbles because of the differences in simulated and real-world sensory information. Shimada’s team worked to overcome this hurdle by adopting a “safety shield” on top of the novel reinforcement learning framework. The safety shield works based on the velocity obstacle concept that says that “the set of all velocities of a robot that will result in a collision with another robot at some moment in time, assuming that the other robot maintains its current velocity.” With this, Shimada’s drones are able to assess which obstacles are high risk for future collision and react accordingly to avoid them.

“Safety is always the most important thing,” said Shimada. “In this case we are looking at making it safer for drones to operate around workers on construction sites, but there are also a lot of dangerous sites where drones can take over jobs to make people safer in those environments too.”

We are looking at making it safer for drones to operate around workers on construction sites, but there are also a lot of dangerous sites where drones can take over jobs to make people safer in those environments too.

Kenji Shimada, Professor, Mechanical Engineering

In the future, Shimada’s lab will work to help drones avoid not just humans on a construction site but moving machinery as well. He hopes to build drones that can react to movement in the same way that flies avoid a swatter—quickly and effortlessly.

This research was published in IEEE Robotics and Automation Letters.