Inverted AI is pleased to announce the initial release of TorchDriveSim [github].  

TorchDriveSim is the first open source lightweight 2D driving simulator built entirely in PyTorch.  It is primarily intended as a training environment for developing autonomous driving algorithms. Its main features are:

1. Fully differentiable execution producing a single computation graph, including state transition (kinematic models) and observation (differentiable rendering) models.
2. First class support for batch processing, down to the lowest level.
3. Support for heterogeneous agent types (vehicles, pedestrians, cyclists, etc.), each with its own kinematic model.
4. Extensible and customizable implementation of kinematic models (unconstrained, unicycle, bicycle, etc.), rendering modes, and rendering backends.
5. Support for extensible traffic control types, including traffic lights.
6. Differentiable implementations of various infraction metrics (collisions, off-road, wrong way).
7. Modular collection of wrappers modifying the simulator's behavior.
8. Ability to ingest any map in Lanelet2 format out of the box.
9. Integration with IAI API for initializing agent states and providing realistic behaviors.

Sign up here for an API key today to start using TorchDriveSim.