The authors reimplemented Atari 2600 games from scratch in JAX, using JIT compilation and vectorization to enable massive parallelism on GPU.
JAX-based Atari environments enable GPU-accelerated training and easy modification
The framework trains PPO on 100 million environment steps in under an hour on a single A100 GPU and exposes object-centric states.
Big Tech
Quentin Delfosse · Raban Emunds · Paul Seitz · Sebastian Wette · Jannis Blüml · Daniel Kirn · +2 more
Technical University Darmstadt · Google Intrinsic AI Research · Hessian Center for Artificial Intelligence (hessian.AI) · Centre for Cognitive Science, Darmstadt · German Research Center for Artificial Intelligence (DFKI)
Research Digest··2 min read
The authors introduce JAXtari, a JAX-native reimplementation of the Arcade Learning Environment that runs entirely on GPU, removing the CPU emulation bottleneck.
§