Yep. It has really cool hiking maps. Especially in Europe.
Plus it is nice to use a non-Google product these days. In the Czech Republic and Slovakia we are lucky to have a company (Seznam.cz) that is able to come up with competitive products and sustain ~50% market share against Google.
I feel the same thing. It's basically optimizing the "gear ratio" to get maximum speed. The wheel isn't centered because it doesn't complete an integer number of revolutions during the simulation.
I'm still scratching my head while trying to figure out what's the mechanism that prevents the solution blowing up to infinity by setting the radius to zero and therefore achieving huge accelerations from a constant torque.
I'm really excited about the prospects of this project. I particularly like the ideas differentiability unlocks once it's a language feature.
For instance, it's possible to write a differentiable rigid body simulator like MuJoCo or Bullet. This has a potential to unlock a new class of reinforcement learning algorithms that can take advantage of the gradients passing directly through in place of very inefficient sampling.
Differentiable simulations in general are the holy grail of optimization and system identification. These ideas have been around since the 70s or so. The bottle neck up until now has been the implementation.
I think a fair comparison would be to AirSim [1] and Carla [3]. These are much more mature projects and are similar to Autodrome in a lot of ways. As far as I know Udacity's simulator is done their class and it's not being actively developed.
+ Both [1] and [3] have much fewer assets (like 3D models of houses, factories, bridges, cars, trucks, and so on) you'd have to buy them on the Unreal/Unity model marketplace and it still wouldn't be enough. Autodrome can take advantage of almost entire Europe and a third of USA at 1:20 scale.
+ Autodrome has a sparse map representation that is really easy to randomly fuzz. I.e. it's easy to shift a segment of the road a little bit and see how the algorithm would react to the fuzzed scenario. I believe this is only way how to achieve robust agents and effectively prevent testing on the training set.
- Biggest disadvantage of Autodrome is a lack of access to in-game dynamic NPCs (like other trucks, cars or pedestrians). As far as I know there's no API for this. Without help (or a lot of very fragile memory hacking) from the developers of the game this feature is very hard to achieve and both [1] and [3] already have it.
Thank you for the detailed response! Is there any thoughts on ground truth segmantic segmentation camera view? Simulated lidar data would be super awesome too.
Both are also not easy like the dynamic NPCs. I think Carla [3] supports both raytraced lidar and segmented rendering now so this is another minus “-“ point.
Plus it is nice to use a non-Google product these days. In the Czech Republic and Slovakia we are lucky to have a company (Seznam.cz) that is able to come up with competitive products and sustain ~50% market share against Google.