Update README with comment about reproducibility.

PiperOrigin-RevId: 795513526
Change-Id: Ibca8fdef1c74a8febad4bcc6b968126cb8140378

Author: btaba

README.md

@@ -89,8 +89,11 @@ python -m rscope
 Get started by installing the library and exploring its features! Found a bug? Report it in the issue tracker. Interested in contributing? If you are a developer with robotics experience, we would love your help—check out the [contribution guidelines](CONTRIBUTING.md) for more details.
 
 ### Reproducibility / GPU Precision Issues
+
 Users with NVIDIA Ampere architecture GPUs (e.g., RTX 30 and 40 series) may experience reproducibility [issues](https://github.com/google-deepmind/mujoco_playground/issues/86) in mujoco_playground due to JAX’s default use of TF32 for matrix multiplications. This lower precision can adversely affect RL training stability. To ensure consistent behavior with systems using full float32 precision (as on Turing GPUs), please run `export JAX_DEFAULT_MATMUL_PRECISION=highest` in your terminal before starting your experiments (or add it to the end of `~/.bashrc`).
 
+To reproduce results using the same exact learning script as used in the paper, run the brax training script which is available [here](https://github.com/google/brax/blob/1ed3be220c9fdc9ef17c5cf80b1fa6ddc4fb34fa/brax/training/learner.py#L1). There are slight differences in results when using the `learning/train_jax_ppo.py` script, see the issue [here](https://github.com/google-deepmind/mujoco_playground/issues/171) for more context.
+
 ## Citation
 
 If you use Playground in your scientific works, please cite it as follows: