Andrej Karpathy (Tesla): CVPR 2021 (see description for better quality version)

For others: Andrej Karpathy is the director of artificial intelligence and Autopilot Vision at Tesla. Was on front page yesterday for a presentation on Tesla's Autopilot / Autonomous features: [VIDEO]

First off, amazing farm-bot project! I am looking forward to reading the details on your site. Thx for the pointers on Tesla. Had not seen the Hydranet stuff. There was a Karpathy talk about 2 weeks back at a CVPR workshop .. he revealed the scale of Tesla's current generation deep learning cluster [1]. It is insane! Despite being in industrial research, I don't foresee ever being able to touch a cluster like that. A lot of our current research involves end-to-end training (some complex stuff with transformers and other networks stitched together). There was a CVPR tutorial on autonomous driving [2], where they pretty much said autonomy 2.0 is all about end-to-end. I've spoken to a few people who actually do commercial autonomy, and they seemed more skeptical on whether end2end is the answer in the near-term. One idea we toy with is to use existing frozen architectures (OpenAI releases some and so do other big players) and do a small bit of fine-tuning. [1] [VIDEO] [2] [LINK]

[VIDEO] In the above video Andrej Karpathy claims that LIDAR required previously compiled maps, the creation of which is not scalable. Do you disagree with that point of view?

I would really like to hear Tesla to defend their decision to not use them. Andrej Karpathy talks some about it in this (it's quite long, but the whole thing is quite interesting): [VIDEO]

Stereo vision is oversold in humans. You can easily drive with one eye. We use a lot of information to estimate depth including relative size of things, as well as “structure from motion” which is the relative movement of things over time. Tesla does estimate depth using structure from motion (and probably other cues). They call it virtual LIDAR in presentations. [VIDEO]

IIRC there was a presentation from Karpathy talking about the challenges with sensor fusion, particular in resolving divergence between e.g. the vision and the radar stack: [VIDEO] My background is in physics, but I find myself having a growing appreciate for the vision-only stack. It's really challenging building a formal understanding of the world that is robust to outliers that are so numerous as navigating in an urban environment. With vision, you have multiple kinds of information that are highly correlated (colour, spatial distribution, depth, etc) that are self-consistent. Whereas, fusing radar with vision, where object responses to radar are highly geometry & material dependent, is a much harder task. I'm really not an expert, so this reads more as an opinion than an experienced view, but I can see the merits in doubling down on vision.