Muse trains the drone in simulation: five flight skills, all passing
A teammate on the project (Muse) ran a drone through a flight simulator and taught it five basic skills: fly a planned square route, circle a fixed point, follow a moving target, run figure-8s and climbs, and hold an exact position on command. All five passed. This is simulation only — no real drone yet — and it's the same see, decide, move idea as our table-top arm, just in the air. The last skill, holding a position from commands sent ten times a second, is the hook a vision 'brain' would later steer with.
What we did
- Flight-path planning: flew a programmed mission — takeoff, a 20 m square at 5 m up, return home, land — and stayed within about 1 m of the planned path (0.27 m average). All four corners hit. PASS.
- Orbit: circled a fixed point at 10 m radius for 70 s, holding the ring to within 0.19 m and height to within 0.15 m. PASS.
- Follow a moving target: tracked a target circling at 15 m for 75 s, settling about 4 m behind it very steadily (0.12 m wobble). PASS.
- Complex maneuvers: a two-loop figure-8 (held to 0.24 m), a fast climb to 8 m then down to 3 m, and a 360° spin in place that drifted only 0.08 m. PASS.
- Precise position holds: obeyed four exact position commands sent ten times a second, holding each within ~0.4 m on average — the same control path a PC vision 'brain' would steer with later. PASS.
What we learned
- Simulation before hardware
- Muse flew all of this in PX4 SITL — a software-in-the-loop flight simulator — with no real drone. Like our arm's MuJoCo twin, the point is to prove the flight skills safely in software before anything real takes off.
- The same see–decide–move loop, in the air
- These five skills (plan a path, orbit, follow, maneuver, hold a position) are the flying version of what our arm does on the table: read the world, decide where to go, move there precisely. The last one — obeying position setpoints at 10 Hz — is the hook a vision brain would steer with.
- Honest-by-design quirks
- The follow test flew higher than the target on purpose (the simulator's default follow height), and that was reported rather than hidden. Good practice: note where a 'pass' includes a known, explainable deviation.
Numbers
| Skills trained / passed | 5 / 5 (PX4 SITL, GPS) |
|---|---|
| Square mission path error | 1.05 m max, 0.27 m mean |
| Orbit radius hold (10 m target) | 10.04 m mean, 0.19 m max error |
| Follow-target settled distance | 4.01 m mean, 0.12 m std |
| Figure-8 path error | 0.24 m max |
| 360° yaw-in-place drift | 0.08 m (turned 345.2°) |
| Position-setpoint hold (4 pts, 10 Hz) | 1.48 m max, 0.39 m mean |
Problems and fixes
Follow mode flew at ~8 m, not the target's 5 m.
Fix: Expected, not a bug: in this simulator, follow mode has a built-in follow height (about 8 m), so it is by design. Logged, not changed.
The next step — the camera-in-the-loop Gazebo vision sim — needs multicast networking the current host blocks.
Fix: Pending: run it on a host where multicast works, or on Ravi's PC.
Next
- Gazebo vision sim on a multicast-capable host (or Ravi's PC).
- GPS-off hover using optical flow.
- A fixed indoor mission, then vision training with person-detection as a safety stop.