As someone completely outside this field, but who likes to keep in touch with progress over the years, I'm often left in awe at the sheer intellectual heft and analytical power that's devoted to self driving. There must several Einsteins quietly at work in these companies around the world. The ability to break down problems of such complexity is almost alien to many people. Can't wait to see which systems come out on top over the next 5-10 years and how that will impact our society and our cities.
Thinking about how a human is driving: we do anticipate other road users' future behavior to adapt our own driving. Such anticipation is done implicitly. We do not predicts other's future in very specific details, such as possible positions at specific future times. We anticipate other in a more general and coarse way, such as the car in front of us may slow down because there is a intersection in front and it is light is already green for a long time (so it may turn red in the near future). This prediction capability is a factor between a good and new driver, we call it experience. Then back to this talk: I don't think anticipate other's behavior is a harder problem, it all depends on what u want to anticipate and how u do it. the CTO is mis-leading the definition. "Prediction" can be anything, predicting other's future behaviors can be understand as estimation of other's intentions. It is the same thing, it related to anticipating future, that's called "prediction"
This is of course great work but one piece of the puzzle that is not as well defined in the original RSS specification is occlusion handling. Has there been any developments on this?
Are 11 cameras input of single 2x board, or multiple ones? And is 10W per chip or board total power? What I mean is - what's total power cost of interpreting 11 cameras with this whole decision making proces? Is it 10W, 20W, or 220W?
10W/EyeQ5 SoC. This number is really irrelevant as the conditions are not specified. Also the number of TOPs is irrelevant if you do not have the right OPs and the right data path. People in the industry know Mobileye well, they do very good engineering work. The main strength of Mobileye is their algorithmic ingenuity and then optimizing the HW to support this algorithms, Tesla basically copied them.
Without a human go make eye contact with and snarl at if they don't let you merge, sorry to say AI cars can't work well. Plus if drivers see car has no human driver they aren't letting it in, there is no chance of human driver yelling so why let this hunk of metal barge in front of me. This is just facts, how a human driver adds that extra persuasion and now taking it away the AI car will be stuck waiting. I think. Plus I walk and bike and make eye contact, AI cars will kill more pedestrians without this.... What a human mess, I'm not happy but this is the situation.. . . . 3rd world people when driver runs over kid at least can beat him up, basic justice, with AI even that is removed, with AI trucks it's 10 people killed at a time, just I have no idea which path is best ... If a kid is killed cause of ai driver I bet the father is gonna assasinate Elon or other CEO, tech nerds are stepping into the physical world and will get some hillbillies angry maybe..... Oh well
As someone completely outside this field, but who likes to keep in touch with progress over the years, I'm often left in awe at the sheer intellectual heft and analytical power that's devoted to self driving. There must several Einsteins quietly at work in these companies around the world. The ability to break down problems of such complexity is almost alien to many people. Can't wait to see which systems come out on top over the next 5-10 years and how that will impact our society and our cities.
Thanks to Mobileye team for publishing of this great video.
Thinking about how a human is driving: we do anticipate other road users' future behavior to adapt our own driving. Such anticipation is done implicitly. We do not predicts other's future in very specific details, such as possible positions at specific future times. We anticipate other in a more general and coarse way, such as the car in front of us may slow down because there is a intersection in front and it is light is already green for a long time (so it may turn red in the near future). This prediction capability is a factor between a good and new driver, we call it experience.
Then back to this talk: I don't think anticipate other's behavior is a harder problem, it all depends on what u want to anticipate and how u do it. the CTO is mis-leading the definition. "Prediction" can be anything, predicting other's future behaviors can be understand as estimation of other's intentions. It is the same thing, it related to anticipating future, that's called "prediction"
weird that he says prediction is not intention. we are predicting their intention and how they will perform the action.
This is of course great work but one piece of the puzzle that is not as well defined in the original RSS specification is occlusion handling. Has there been any developments on this?
may I know how to guarantee the accuracy of intention model?
a long winded master of ceramony
Are 11 cameras input of single 2x board, or multiple ones? And is 10W per chip or board total power? What I mean is - what's total power cost of interpreting 11 cameras with this whole decision making proces? Is it 10W, 20W, or 220W?
10W/EyeQ5 SoC. This number is really irrelevant as the conditions are not specified. Also the number of TOPs is irrelevant if you do not have the right OPs and the right data path. People in the industry know Mobileye well, they do very good engineering work. The main strength of Mobileye is their algorithmic ingenuity and then optimizing the HW to support this algorithms, Tesla basically copied them.
@@dustybeardall what design corner and what workload ?
🍟
Without a human go make eye contact with and snarl at if they don't let you merge, sorry to say AI cars can't work well. Plus if drivers see car has no human driver they aren't letting it in, there is no chance of human driver yelling so why let this hunk of metal barge in front of me. This is just facts, how a human driver adds that extra persuasion and now taking it away the AI car will be stuck waiting. I think. Plus I walk and bike and make eye contact, AI cars will kill more pedestrians without this.... What a human mess, I'm not happy but this is the situation.. . . . 3rd world people when driver runs over kid at least can beat him up, basic justice, with AI even that is removed, with AI trucks it's 10 people killed at a time, just I have no idea which path is best ... If a kid is killed cause of ai driver I bet the father is gonna assasinate Elon or other CEO, tech nerds are stepping into the physical world and will get some hillbillies angry maybe..... Oh well