Magic Leap One Ships: Reflections on Launch Day
We shipped a mixed reality headset. Reflections on what it means, what we got right, what we got wrong, and what comes next.
Magic Leap One Creator Edition is shipping today. After two and a half years, the device is real and going to customers.
What We Shipped
The perception system in every Magic Leap One includes:
- 6DoF head tracking: Visual-inertial odometry at 60Hz, under 1mm accuracy indoors
- Spatial mapping: Real-time 3D mesh generation
- Plane detection: Horizontal and vertical surfaces for content placement
- Eye tracking: Gaze vector for foveated rendering and interaction
- Image tracking: Fiducial markers for spatial anchors
Every one of those was a research problem when I started. They are production features now.
What Went Right
The synthetic data investment paid off. Our eye tracking and hand tracking models lean heavily on procedural data generation, and we could never have collected enough real data to match it. Calibration held up too: the production calibration system works, yield is above target, and quality is consistent across units. In April I wrote that hitting the yield target would be worth more than any other optimization available to us; yield is above target, and I stand by the ranking. The sensor architecture calls look good in hindsight - ToF depth and global shutter tracking cameras were the right choices, so we are not fighting sensor limitations on top of everything else.
The team held together through a brutal final push. Nobody burned out catastrophically.
What Went Wrong
Outdoor performance, first. SLAM doesn't work reliably outside. We knew that going in and hoped to solve it anyway, and we didn't; it ships as a known limitation. Power consumption came in higher than spec, which means battery life is shorter than we wanted and thermal throttling kicks in sooner than ideal. Gesture recognition was one of July's hard cuts, so users get basic hand tracking without the natural interaction we envisioned. Two of the four priorities I set for 2018 are in this paragraph: we never got back inside the power budget, and hand tracking shipped as half an input modality. And there is a polish gap: many edge cases have rough edges. The core works, but it doesn't feel finished the way Apple products do.
Market Reality
The reviews are... mixed. The technology is impressive, the price ($2,295) is high, the content ecosystem is thin, and the form factor is chunky. We built what we could build. The market will decide if it's enough.
What I Learned
Shipping is a muscle. You only learn to ship by shipping, and all the planning and process matter less than the experience of actually getting something out the door. Hardware is unforgiving in a way software never taught me: software bugs can be patched, but hardware ships with its limitations baked in, and the decisions we made in 2016 determined what was possible in 2018.
The device itself was never the point. The goal is enabling experiences that weren't possible before, and for that the device is necessary but not sufficient.
What's Next
For the product, the near-term list is software updates to improve tracking, content partnerships to fill the experience gap, and iteration toward V2 with the lessons we just paid for. For me, the shift is from V1 firefighting to V2 architecture, scaling the synthetic data platform, and working on the next generation of perception capabilities.