Where it learns

Before the aircraft flies, it flies here. What is in the simulator, what we ask of it, and what the aircraft on the other side of it looks like.

What is in the world

Two scenes so far. A race circuit with a car that laps it, and a cluttered warehouse floor with a four-legged robot walking through it. Both are simple on purpose. The point is not a pretty world, it is a target that moves in ways a waypoint list cannot predict.

The aircraft in the simulator is our compact drone, built from the same CAD model as the hardware, with the same sensors in the same places. When the policy moves to real hardware it should not notice.

02The circuit, from above.

The task

One tool: follow. The aircraft is given a target and asked to keep it in frame at a standoff distance, without hitting anything. It is not given a path. The flight is a learned policy, trained by reinforcement in the simulator, and the only thing it is told is what to watch.

Follow is deliberately small. It is the piece most real jobs decompose into, and it is the one where spatial context beats a waypoint list every time: the target moves and the waypoints do not.

  • InputWhere the target is, where the aircraft is
  • OutputWhere to fly next
  • Scored onTarget in frame, standoff held, nothing hit
  • PathNot scripted

The aircraft

A compact quadrotor designed to develop against, because the interface is only real when it is attached to sensors that exist. It is small on purpose. Small is cheap to crash.

  • Flight controllerXIAO ESP32-S3 on a carrier
  • Propulsion4x brushless, 31 mm props
  • Inertial, baro, magOn carrier
  • Depth, front and rearVL53L5CX
  • Downward ranging, optical flowToF + flow
  • PositioningUWB daughterboard
  • Power1S 480 mAh, V and I sensing
03The aircraft model tracking a quadruped.

What the simulator is for

Three things. It is where policies are trained, because crashing a simulated drone ten thousand times is free. It is where the checks are exercised, because a geofence should be argued with in software long before it is argued with in the air. And it is where the interface gets its first server: the simulated aircraft speaks the same tools and exposes the same sensors as the real one will.

The clips are on the videos page. The reason we are doing any of this is the thesis.

Training details and sim-to-real results are not on this page. If you want them, ask.