A drone that learns in a simulator first: five flight tests, all passed
Our robot arm learned its first skills in a simulator before it ran on the desk. We are doing the same with a drone. The aim is a drone that flies by what its camera sees, and stops when a person is in the way.
Five tests in a simulator
On November 7 we ran five flight tests in a simulator, with PX4, the open-source flight software used on many drones. No real drone has flown yet. This is all in software, so we can prove each skill safely before anything takes off. All five passed.
| Test | What the drone did | Result |
|---|---|---|
| A planned path | Took off, flew a 20 m square at 5 m, came home and landed | 0.27 m off the planned line on average, 1.05 m at most |
| An orbit | Circled a fixed point at a 10 m radius for 70 seconds | Radius 10.04 m on average, never more than 0.19 m off |
| Follow a moving target | Followed a target that circled at a 15 m radius, for 75 seconds | Settled about 4 m behind it and held steady |
| Hard moves | A figure 8, a fast climb to 8 m, a drop to 3 m and a turn in place | The figure 8 stayed within 0.24 m of the plan |
| Commands from our code | Our Python script sent four points to fly to, 10 times a second | Held each point within 0.39 m on average |
Why the last test matters most
The last test is the path the vision brain will use. A computer works out where the drone should go and sends it a point to fly to, many times a second. Now we know the drone goes where our code tells it to.
What comes next
- A simulator with a camera, so the drone can practise seeing.
- Holding a hover without GPS, with a camera that watches the floor (optical flow).
- A fixed indoor route.
- Then flying by sight, with person detection as a safety stop.
Every lab day goes into the public robotics journal. Follow it at robotics.autosterea.com/journal.
