blog
blog
Out of context: Reply #68563
- Started
- Last post
- 75,615 Responses
- shapesalad0
https://www.bbc.com/news/technol…
We learn from mistakes.
Current approach for self driving cars is to fit it with sensors and program it to read the road, recognise this and that and at all costs avoid a crash.
But how about fitting a car with a massive thick bumper all around it, giving the deep learning AI visual and 3d laser scan data, and setting it to go from A > B. Each time it bumps into something, the AI reviews the proceeding data and learns what to do to avoid it. Eventually the AI will have learnt how to drive and avoid crashes. You just need a test course to practice it on. Once it's 90% there, get it on real roads with a human intervening when it makes a wrong move.
- You think they let AI vehicles out onto the streets that are anywhere near '90% there'?Nairn
- They kind of already do what you suggets by having the driver behind the wheel. Every time the driver's input is invoked, it's logged and fed back.Nairn
- That's the supposed strength of Tesla's path - they have n thousand data points, constantly feeding back.Nairn
- Waymo has 5 million miles of road testing, and 5 *billion* miles of simulation testing under its hood, so to speak, and still not 'production ready'.Nairn
- https://www.theverge…shapesalad
- All Teslas constantly monitor driving conditions to improve all car AI in the fleet, regularly. Exactly like this.monospaced
- squishy foam cars?robotron3k
- ah...shapesalad
- maybe they are trying that. not a bad ideadocpoz