Reproduciability of AirSim-trained networks in real autonomous driving situations

microsoft / AirSim

Open source simulator for autonomous vehicles built on Unreal Engine / Unity, from Microsoft AI & Research

https://microsoft.github.io/AirSim/

Other

16.55k stars 4.61k forks source link

Reproduciability of AirSim-trained networks in real autonomous driving situations #1471

Closed eds89 closed 5 years ago

eds89 commented 6 years ago

Hello guys,

Have you ever seen an assessment of how well neural networks trained in MS AirSim's simulated world work in real autonomous driving situations? I'd love to see a detailed scientific study on this. Some top researchers at my university disregard simulation technology altogether.

DavidLSmyth commented 6 years ago

In relation to AirSim, I saw these at IJCAI this year: https://www.ijcai.org/proceedings/2018/0678.pdf https://www.ijcai.org/proceedings/2018/0847.pdf

It seems to me like there's plenty of interest in using simulations and transfer learning to train systems due to the fact that simulations offer a ground truth, but I don't know of any detailed studies outlining a methodology evaluating their effectiveness. I'd also be interested if anyone has any quantitative results or methodology.

ashkapoor commented 5 years ago

See: https://github.com/Microsoft/AirSim/wiki/technion