Craig Quiter
Craig Quiter
Hey Andrej! Cool repo :+1: This is a small change that effectually computes `model(x - 0.5)` in `visualize.ipynb`. Otherwise `model(x) - 0.5` looks like this  After the change we...
Per [this comment](https://github.com/deepdrive/deepdrive/issues/12#issuecomment-386829928) in #12
We should enable wide variation in camera output to encourage [domain randomization](https://arxiv.org/abs/1703.06907) within agents trained on camera data. Like start position and other [domain-randomization features](https://github.com/deepdrive/deepdrive-sim/labels/domain-randomization), this should be controlled by...
Gear ratio, engine dampening, differential bias, torque curve, 4wd vs 2wd, etc...
Behavior of traffic should be driven by random seed passed from client in `train` mode. In `test` mode random seed should be generated by a cryptographically secure pseudo-random number generator...
Basically reimplement [this](https://github.com/deepdrive/deepdrive/blob/698d9d89f026d34892463cb9f5d58bcfc10467c3/agents/dagger/agent.py#L417-L433) in C++ and put it behind the global random seed - changing the camera(s) after every reset.
Random obstacles should be placed in the road to make the current track more stochastic. The in-game PID should route around these as well.
The last collision time should be returned via the `PyCaptureSnapshotObject` for the ego (agent) vehicle. In `test` mode this should end the episode and require a reset from the client...