Great summery of our multi-agent quadrotor racing project in @jackclarkSF's ImportAI newsletter [1]!
While a massive success for racing, the project opened more questions than it answered, going far beyond technical aspects.
For now, additional technical bits: Multiagent learning really became key for us to transfer from simulation to human behaviours. It directly mirrors the role of targeted domain randomisation for transferring to new visual or dynamic aspects #sim2real
The perceiver choice actually nicely reflects recent successes for offline RL from previous colleagues @GoogleDeepMind [2]