Talk Date and Time: September 12, 2023 at 1:00 pm - 1:45 pm EST followed by 15 minutes of Q&A on Google Meet
Topic: Multi-Agent Simulation and Learning in TorchRL
Abstract:
In this talk, we will discuss how multi-agent simulation and learning can be performed in the TorchRL library. In particular, we will focus on showcasing TorchRL's MARL API through a series of examples and demos from the multi-robot systems domain.
The talk will begin by introducing the VectorizedMultiAgentSimulator (VMAS), a vectorized simulator comprised of a PyTorch physics engine and a collection of multi-robot tasks. It will then focus on discussing how this simulator is integrated in TorchRL training library to benefit from on-device batched simulation and training as well as illustrating the general API for integrating any MARL environment/game in the library.
Lastly, it will present an application of the components presented through a live demo of a full multi-agent training pipeline for a multi-robot navigation task.
Bio:
Matteo is a PhD student in the ProrokLab at the University of Cambridge and an Intern at PyTorch Meta. He holds an MPhil in Advanced Computer Science from the University of Cambridge and a BEng in Computer Engineering from Politecnico di Milano. His research focuses on studying resilience and heterogeneity in multi-agent and multi-robot systems. For his research, he employs techniques from the fields of Multi-Agent Reinforcement Learning, Robotics, and Graph Neural Networks. Further research and information about him can be found at https://matteobettini.com/