Movers already fly, so why Magflow?
Even today, low-level control is exclusively domain of handcrafting and engineering
These approaches work, but are conservative
Movers and tiles are co-designed to allow handcrafting (this enabled real-time control), but...
→ Controllers inherit constraints from co-design, e.g. yaw rotation is hard and supports only small weights
→ Only movers with complex magnet topologies (Halbach arrays) can be controller
Magflow enables:
First unconstrained 360° yaw rotation at any point (of a single-tile single-mover system)
First levitation of a simple bar magnet
First 360° rotation about a horizontal axis
Many more in future
Features:
Simulation replicates a real system used in industry
One computational graph (JAX, GPU) for rollout (neural controller + physics)
Parallelized rollouts of multiple domain randomized worlds
4khz simulation/real system
Trained neural controllers transfer zero-shot to hardware
Rollout of a learned neural controller which roates the bar magnet in yaw 360°
360° yaw rotation away from tile center with APM4330 mover
360° yaw rotation away from tile center with bar magnet
360° roll rotation with bar magnet
at least 4x weight capacity than prior neural controller