Date: 5th July 2025
arXiv link
Key Points:
- PyTorch version of IMPALA
- Attempts to balance high performance low level code with ease of implementation of PyThon.
- E.g. batching is done by C++.
- Presents a high performance server set up:
- Many environment servers: send observations and rewards
- Client server: does neural network processing. Receives observations and sends back actions.
- One learner server: learns.