I wonder if you have an obstacle directly in front of the car and half of the people drive left and half decide to drive right, would the consensus be the middle?
I think there could conceivably be something like that, but on real roads there would be a clear split between the part of the lane where you choose to pass into the fast lane instead. Over time, if there are proper blindspot cameras and radars, then you also have those inputs to help make the decision. I think it's rare that people choose a direction for _no reason whatsoever_ .
This is a really interesting/scary question. I *think* what would happen is the model would initially output moving in one direction, and then subsequent decisions on the future of the path would remain biased to that one directional decision, i.e. the initial choice to go left is continuously reinforced by the model which says "once a human started turning left, they continued to turn more and more left so I will do so too". But I'm a total AI amateur.
Tesla is probably doing the same thing with a lot more data and a lot more resources(might not be a good thing tbh). Plus, the classic approach can provide a nice human understandable intermediate output, which might never be complete, but does provide a nice UI(Look at this crazy feature!). I think the main difference here is Tesla only can collect data for Tesla cars, which means the model they train will likely only able to work on Tesla cars, but comma will work on all different brands of cars.
One big problem seems like it’s system really depends on the hardware.. like if in a new version they changed the hardware, then they have to create a completely new version of the training..
@@HimanshuGhadigaonkar You mean cameras? You really only need two cameras(one facing forward and one backward) anyway and I don't think they will change the form factor much for these cameras.
to claim that any current model is superhuman would be disingenuous. setting better than the average human ("superhuman") as your goal seems reasonable to me :). i mean that _is_ what they're going for.
A perfect driver is an agent that goes fro A to B through chaos without colliding into any other object. I would train a perfect collision avoider but it would work better if the car could move like an agar.io player. How do you train the perfect collision avoider? I've been seeing this idea about learning from how humans drive and it is genius, sadly humans are really bad drivers though the only drivers in the planet.
On German cities and autobahn there are quite some instructions written (which days you can or cannot access the road, what speed to go under what conditions, when and where you can park). Good luck end to ending that
It seems like a relatively simple problem at first glance but must be endlessly complicated when you dive into it. For instance, taking into account the characteristics of every vehicle type, or every instance of each car with its own calibration, not to mention implementing navigation into the system... there are so many factors that can make the system break when it should clearly NEVER break. Also, I cannot help but feel a little bit of involuntary arrogance when they compare human brains to artificial neural nets. Human beings are amazing. We can learn driving in just a few sessions, with so very little "input data", compared to this. It's clearly two very different things. I wish comma a lot of success, they seem to make great progress though, and seem have the best approach considering the current state of tech. However, I fear we might never be able to achieve level 5 without AGI, and that doesn't seem to be on the near horizon...
Maybe. I still feel like even though it can work in 99% of the time, the 1% missing might only be attainable through actual reasoning and "common sense". What I'm talking about is a fully autonomous, with zero human intervention driving. I guess we'll see in the next years. Technology is going to evolve further too.
This guy knows his stuff... Also happy birthday! 🎂🎉🎊
This just in: Comma takes a break from self driving cars to develop a self clicking clicker
Yeah. Thankfully there are no moving parts on the comma to stick or break. Although, what happens if the comma has a kernel panic? Wake up, driver
This is super impressive. Seems way more creative than Tesla's approach. Super excited for all the progress in this space.
yeah but I think Tesla wants to get to lvl 5 right away and comma just wants like lvl 2 for now
What about obstacle avoidance? it seems you would have to collect a huge amount of obstacle course type data to give the model enough context.
Simulator
Sad that they've been having technical difficulties with the clicker setup. Great talks though!
I’m impressed by the amount of water that he needed to drink.
They seriously gotta figure out their clicker situation next time lol Pretty interesting, awesome presentations tho!
the one guy from the q&a works on the soviet self driving car haha
Comma Con: nice
Tesla AI Day: Poggers
I wonder if you have an obstacle directly in front of the car and half of the people drive left and half decide to drive right, would the consensus be the middle?
I think there could conceivably be something like that, but on real roads there would be a clear split between the part of the lane where you choose to pass into the fast lane instead. Over time, if there are proper blindspot cameras and radars, then you also have those inputs to help make the decision. I think it's rare that people choose a direction for _no reason whatsoever_ .
This is a really interesting/scary question. I *think* what would happen is the model would initially output moving in one direction, and then subsequent decisions on the future of the path would remain biased to that one directional decision, i.e. the initial choice to go left is continuously reinforced by the model which says "once a human started turning left, they continued to turn more and more left so I will do so too". But I'm a total AI amateur.
Very interesting! I feel like this approach is even better and more quickly scalable than Tesla's
Tesla is probably doing the same thing with a lot more data and a lot more resources(might not be a good thing tbh). Plus, the classic approach can provide a nice human understandable intermediate output, which might never be complete, but does provide a nice UI(Look at this crazy feature!). I think the main difference here is Tesla only can collect data for Tesla cars, which means the model they train will likely only able to work on Tesla cars, but comma will work on all different brands of cars.
One big problem seems like it’s system really depends on the hardware.. like if in a new version they changed the hardware, then they have to create a completely new version of the training..
@@HimanshuGhadigaonkar You mean cameras? You really only need two cameras(one facing forward and one backward) anyway and I don't think they will change the form factor much for these cameras.
eating cake with your hands is never a mistake harald.
Would love to see OpenPilot 0.9 and if there will be any features for obstacle detection such as road debris, cattle, and deer
Brilliant! I Love both the focus and results
Good evening I have some ideas to implement in my startup
Still remember George "bash" other autonomous driving developers using the keyword "superhuman" everywhere in their presentation.
to claim that any current model is superhuman would be disingenuous.
setting better than the average human ("superhuman") as your goal seems reasonable to me :). i mean that _is_ what they're going for.
Dear Heorge Hotz and the Comma Team, is the Comma Logo derived from the Sharingan? Please answer me george, it does not let me sleep at night!
Please ship longitudinal as a toggle asap!!
A perfect driver is an agent that goes fro A to B through chaos without colliding into any other object. I would train a perfect collision avoider but it would work better if the car could move like an agar.io player. How do you train the perfect collision avoider? I've been seeing this idea about learning from how humans drive and it is genius, sadly humans are really bad drivers though the only drivers in the planet.
@@JayDee-b5u Why not first train a general avoider, and then a car driver that is a subset of a general avoider.
I ask your new software comma three can pass from front cars or make drifting ? if not then there is some codes not complete
On German cities and autobahn there are quite some instructions written (which days you can or cannot access the road, what speed to go under what conditions, when and where you can park). Good luck end to ending that
that is not something openpilot will attempt to solve. the scenarios you're describing are when "not" to drive entirely.
It seems like a relatively simple problem at first glance but must be endlessly complicated when you dive into it. For instance, taking into account the characteristics of every vehicle type, or every instance of each car with its own calibration, not to mention implementing navigation into the system... there are so many factors that can make the system break when it should clearly NEVER break.
Also, I cannot help but feel a little bit of involuntary arrogance when they compare human brains to artificial neural nets. Human beings are amazing. We can learn driving in just a few sessions, with so very little "input data", compared to this. It's clearly two very different things. I wish comma a lot of success, they seem to make great progress though, and seem have the best approach considering the current state of tech. However, I fear we might never be able to achieve level 5 without AGI, and that doesn't seem to be on the near horizon...
Maybe. I still feel like even though it can work in 99% of the time, the 1% missing might only be attainable through actual reasoning and "common sense". What I'm talking about is a fully autonomous, with zero human intervention driving.
I guess we'll see in the next years. Technology is going to evolve further too.
if i can sleep all the way, then who cares if it's super slow?
The emperor has no clothes.
👍🏼
Can't even get the dang poewrpoint to work. I'm not sure I can trust self driving cars :joy:
No such thing as "the average human". The "median human" crashes a lot less than the mean (how much, I don't know, but it could be 2-10x)
I'm sorry... you're a tech company that can't get a powerpoint right, and we should feel good with you piloting 2-ton death machines?
Hello Harald from research.
Please don't make the Machine, thank you.