This week was the unglamorous-but-necessary work of making the simulator behave like a real drone, not a Gazebo-default toy. Most of the time was spent on motor dynamics — getting the simulated iris-claudedrone to accelerate, decelerate, and lose altitude the way a real ~600 g indoor quad with brushed 1404 motors actually does.
What landed
- Custom motor physics plugin. The default Gazebo
LiftDragPluginis acceptable for outdoor scale drones but produces unrealistically snappy responses on a small indoor quad. Replaced with a tuned plugin that models motor spin-up time constants, propeller efficiency curves, and a more honest battery-sag model. Details: gazebo/plugins/custom-motor-physics. - URDF refinements. Center-of-mass alignment, moment-of-inertia matrix updates based on the planned BOM, and prop-mount stiffness — all of which subtly affect how the drone responds to step inputs. Details: gazebo/models/drone-chassis-urdf.
- Step-response calibration. Manually flying the simulated drone with the same
MANUAL_THROTTLEsetpoints we’d use on the real one, comparing the altitude/velocity traces against published step responses for similar-class drones. Not perfect, but in the ballpark.
Why this matters
Sim-to-real transfer is the entire deliverable for the RL track. A policy trained in a Gazebo world that doesn’t match the real drone won’t transfer. The community wisdom is that physics fidelity matters more than visual fidelity — the policy doesn’t see the rendered scene, it sees rangefinder distances and IMU readings, and those depend on how the drone moves through space.
We don’t yet know how good our physics match is — that question won’t get answered until we have a real drone to compare against, which is on the H2 2026 roadmap. What we can say is that the simulated drone now passes the visual sanity check: it doesn’t accelerate like a fighter jet on a 30%-throttle command, and it doesn’t drift like it’s in zero-G after a setpoint change.
Open questions we deferred
- Propeller wash interaction. Two adjacent drones (or a drone near a wall) experience aerodynamic interaction that our current physics doesn’t model. For single-drone indoor work that’s fine; for multi-drone work it’s a gap. Deferred.
- Battery voltage sag under sustained load. The current model is a linear discharge curve; real LiPos sag nonlinearly under high-current draw. Acceptable for now since our episode lengths are short; will revisit before real-drone deployment.
Where this leads
- The bigger Gazebo writeup: Gazebo simulation environment
- ArduCopter parameters tuned for this physics setup: arducopter-params-indoor
- Next month picks up with the heavier RL experiment sequence — latest RL dev-log