NAVIX

NAVIX achieves over 1000x speedup compared to original MiniGrid, reducing week-long experiments to minutes, as demonstrated in benchmarks showing 670 million steps per second.

Leverages XLA compilation to run on CPU, GPU, and TPU, enabling scalable performance across different accelerators without code changes.

Supports autograd through environment steps, allowing backpropagation for novel research like learned world models, as shown in the examples.

Enables batched execution of thousands of agents in parallel, facilitating hyperparameter tuning at scale and reducing experimental noise.