Reproducing paper results

[HrishikeshVish]

Could you please share the hyperparameters used for the experiments in the paper?

[michael-psenka]

It should just be the default parameters here: https://github.com/michael-psenka/grasp/blob/main/conf/planner/grasp.yaml

Note the paper's results were first-to-success: i.e. 500 runs were made in parallel, then each run was exited as a success once it reached the goal at any point, and success time was measured there. GRASP fundamentally is faster at finding goals, but at its most basic implementation the reshaped gradient structure can sometimes destabilize at a goal-reaching plan. There are some straightforward ways to amend this, e.g. a linearly annealing schedule in noise/goal loss that I implemented in the le-wm and jepa-wm forks (see README, "More Environments and World Models"), but not currently implemented here.

Hope this helps! Let me know if you have any more questions.