Recently I heard about an approach where they train a model to behave as a Lidar, using only camera inputs. Sounds like a cool, modular approach, but not sure about safety.
You can train those models all you want, the sensors still have some physical limitations. A regular camera operating in the visible spectrum won't be able to penetrate dense fog, where lidar or radar may still work.
Meh, i believe Tesla will be wrong on this one. Yes, they will achieve good autonomous cars without lidar and radar, but someone else will build a slightly safer autonomous driving experience with those added modalities. The costs will plummet as the sensors become ubiquitous and eventually they will be required by legislation.
Yup, I strongly agree with you too. Infact In fact I mentioned that in my interview for the Tesla Autopilot team lol I guess that’s why they didn’t move me forward in the process.
Tesla FSD used to do this with occupancy networks [0]. It's basically a "software lidar". Afaik they don't use it any more in their newest FSD stack, expect for visualizations maybe.
reply