SLAM improvements from 2 years of Crystal Light

[00:00:10]Hello everyone, here's Steve. Back in 2024, our slam algorithm was honestly, just okay. But now, its performance is just as shown at the beginning of the video, more precise and stable. In this video, we will walk you through what our engineers have been optimizing over the past 2 years. And also, we'll lead you to have a look what is still coming for Pimax slam tracking at the end of the video. Our early Crystal light units, the most reported issue was jitter. A quick run, a slight movement, even standing completely still, the image would tremble. I know it, that really breaks immersion. We approached this from two directions.

[00:00:49]First, we introduced line feature tracking. Street edges, wall, door frames, give the algorithm more visual anchors. More reference points means less jitter. Second, we tackle micro jitter stationary states. This isn't the kind of problem you'll fix in single update.

[00:01:08]>> [music] >> It takes sustained effort over time.

[00:01:11]We've been steadily working it down for nearly 2 years.

[00:01:15]Second issue, drift. [music] Inside-out tracking relies on camera continuously recognizing environmental features. When lighting changes or feature points are lost, position estimation gradually drifts. Extended session, bright light source, the visual space slowly drift.

[00:01:32]At best, it's uncomfortable. At worst, you need to relocate. In 2024, we made a major optimization pass on this. After that release, drift and tilt probability dropped a lot. Long session stability reached a level we were satisfied with >> [music] >> for the first time. For bright light scenario, we're building independent exposure adaptation. Windows, direct lighting, they're no longer high-risk triggers for drift. [music] The next is relocalization. Previously, every time you powered on, you had to redraw your boundary. And if tracking was lost, recovery was slow. In July 2024, we shipped offline relocalization. Power off and power on, the headset recognized your space and reset your boundary automatically. No need to do it by yourself. You just go. Later updates keep improving accuracy, heat drift, and positional offsite after restart have been gradually reduced, and relocalization speed is significantly faster than it was. Fourth, controller tracking, specifically aiming. Early versions have clear shortcomings in shooter scenario. [music] Jitter, occlusion, and controller lag were all issues that directly affected the experience. These are the moments shooter players care about most, and we knew that we needed to do something to fix it. We went back to the detection logic and rebuilt it. The way that algorithm find each controller's position was redesigned, which fixed a key problem where our crossed controllers would confuse the tracking.

[00:03:03]We also made the system more stable under occlusion and bright lighting. So, losing sight of a controller briefly won't cause it to drift. For aiming, we did a dedicated pass on stability when the arm is fully extended, and the speed at which a controller is picked up again from leaving the camera's view has been greatly improved. I know we still have a lot to do, but yeah, we're working on it. This is 2 year of iteration and growth in Pimax SLAM tracking. None of this happens without your feedback.

[00:03:33]Every report, every suggestion, it all made a difference. All right, there's one more thing we've been working on, [music] hand tracking powered by SLAM.

[00:03:42]We have actually mentioned this feature all along in our previous videos, but this time we have a really made some breakthrough. The tracking is now lower latency, more accurate, and more stable.

[00:03:53]And we build on that to add a bunch of new practical gesture controls. We will share more details in the future video.

[00:04:00]We can share it with you as soon as possible. That is all for today's video.

[00:04:03]Thank you to everyone who's been with us along the way. See you next time.