[bitL]

All open source state-of-art SLAM libraries fail on real-world data, often the same scene that was processed once successfully fails in another pass. The problem is with the computer vision algorithms themselves, the "classical" ones are super susceptible to per-scene constants tuning and randomization effect; the only hope IMO is Deep Learning at the moment, but that requires massive computational capability for real-time inference.

[stanlarroque]

Yes, the best SLAM we have today based on visual inertial data (Camera+IMU) uses 500mW of power. It uses DSPs and ASICs. Also, the custom silicon team from MSFT did a great job for their SLAM and display engine, it's built on ASICs in what's called the HPU for Hololens: https://youtu.be/u0eBd2m_wEs?t=1641