We release several significant additions, such as support for:
-
TorchRL and BenchMARL: several SoA MARL algorithms that support heterogeneous groups too.
-
Success/failure criteria: a significant number of navigation tasks (e.g., "move safely to a location") require/offer binary outcome.
-
Global state to increase training stability in MARL tasks.
-
Learning to navigation and communicate is a very interesting problem, which we can now tackle using
navground-learning. We added a (quite long) tutorials analyzing a prototypical task around navigation+communication.
Full Changelog: 0.1.1...0.2.0