I wish my car was supported... I think that it isn't because the steering motor is too weak (Mazda CX-5).
sidenote: if you navigate to a non-existent link on their domain, all links stop working... ie: visit https://comma.ai/404 and click on any link at the top
Direct link to the YouTube video if you don't care about the blog post with technical details: https://youtu.be/SUIZYzxtMQs
This is really impressive and demonstrates capability comparable to Tesla FSD.
I've been a Comma enthusiast for years now, and it's so impressive how they've been able to consistently stay 1-2 years behind Tesla in terms of capability, all with a tiny team. I think they have only 15 engineers or so.
I wouldn't be surprised if they totally catch up to Tesla, unless Tesla changes their approach.
The main difference in their approaches is that Tesla is trying to use vision to build a full internal 3D model of the world. Detect every cone and load a cone into their virtual model, detect every lane line, detect every road sign. Load them all into the virtual model. Then do planning in virtual model space.
Comma is just committed to full end-to-end planning. "In this situation, what would a human do?". They don't try to build this intermediate virtual representation.
The advantage of Tesla's approach is it's more introspectable and debuggable. You can see why it made the choices it made. The downside of this approach is that doing planning in virtual space with a path planning algorithm results in a very mechanical driving style that does not feel super natural to the passengers.
The advantage of Comma's approach is it feels much more natural -- after all, it's doing what it predicts a human would do in each situation. Another advantage is their data engine is just a ton simpler. This is partially how they get away with having so few employees and making such a good product. The disadvantage of the "pixels to pedals" end-to-end approach is it's hard to tell why a model did something. "Because that's what the neural network thought a human would do" is not a super useful answer.
It is possible you could build an auxiliary neural network that took the hidden activations of the main neural network as input and tried to extract "thoughts" from it, e.g. "this neuron always turns on when there is an obstacle in the lower right field of view".
Something I love about Comma's approach is it allows them to just add more sensors and data sources over time. Like how to add the full navigation following part of self driving? Just stick a convolutional neural network to look at the image of the nav screen, as if it's just another camera looking at something.
3 comments
[ 2.8 ms ] story [ 20.2 ms ] threadsidenote: if you navigate to a non-existent link on their domain, all links stop working... ie: visit https://comma.ai/404 and click on any link at the top
This is really impressive and demonstrates capability comparable to Tesla FSD.
I've been a Comma enthusiast for years now, and it's so impressive how they've been able to consistently stay 1-2 years behind Tesla in terms of capability, all with a tiny team. I think they have only 15 engineers or so.
I wouldn't be surprised if they totally catch up to Tesla, unless Tesla changes their approach.
The main difference in their approaches is that Tesla is trying to use vision to build a full internal 3D model of the world. Detect every cone and load a cone into their virtual model, detect every lane line, detect every road sign. Load them all into the virtual model. Then do planning in virtual model space.
Comma is just committed to full end-to-end planning. "In this situation, what would a human do?". They don't try to build this intermediate virtual representation.
The advantage of Tesla's approach is it's more introspectable and debuggable. You can see why it made the choices it made. The downside of this approach is that doing planning in virtual space with a path planning algorithm results in a very mechanical driving style that does not feel super natural to the passengers.
The advantage of Comma's approach is it feels much more natural -- after all, it's doing what it predicts a human would do in each situation. Another advantage is their data engine is just a ton simpler. This is partially how they get away with having so few employees and making such a good product. The disadvantage of the "pixels to pedals" end-to-end approach is it's hard to tell why a model did something. "Because that's what the neural network thought a human would do" is not a super useful answer.
It is possible you could build an auxiliary neural network that took the hidden activations of the main neural network as input and tried to extract "thoughts" from it, e.g. "this neuron always turns on when there is an obstacle in the lower right field of view".
Something I love about Comma's approach is it allows them to just add more sensors and data sources over time. Like how to add the full navigation following part of self driving? Just stick a convolutional neural network to look at the image of the nav screen, as if it's just another camera looking at something.