The Flying Machine Arena (FMA) stopped operating at the end of 2019.
Previously the FMA was a movable space devoted to autonomous flight. Measuring up to 10 x 10 x 10 meters, it consisted of a high-precision motion capture system, a wireless communication network, and custom software executing sophisticated algorithms for estimation and control.
The motion capture system could locate multiple objects in the space at rates exceeding 200 frames per second. While this may seem extremely fast, the objects in the space could move at speeds in excess of 10 m/s, resulting in displacements of over 5 cm between successive snapshots. This information was fused with other data and models of the system dynamics to predict the state of the objects into the future.
The system used this knowledge to determine which commands the vehicles should execute next in order to achieve their desired behavior, such as performing high-speed flips, balancing objects, building structures, or engaging in a game of paddle-ball. Then, via wireless links, the system sent the commands to the vehicles, which executed them with the aid of on-board computers and sensors such as rate gyros and accelerometers.
Although various objects could fly in the FMA, the machine of choice was the quadrocopter due to its agility, its mechanical simplicity and robustness, and its ability to hover. Furthermore, the quadrocopter is a great platform for research in adaptation and learning: it has well understood, low order first-principle models near hover, but is difficult to characterize when performing high-speed maneuvers due to complex aerodynamic effects. We were able to cope with the difficult to model effects with algorithms that use first-principle models to roughly determine what a vehicle should do to perform a given task, and then learn and adapt based on flight data.