I run an experiment with perfect GPS & Compass but noisy actuations and input sensors. Trained till 970 millions, val performance: success: 0.9386(Harsh's 0.97), spl: 0.6536(Harsh's 0.71), softspl: 0.6448
while train performance on train: distance_to_goal: 0.110 reward: 9.105 softspl: 0.774 spl: 0.790 success: 1.000.
Added training of the agent with GT localization:
I run an experiment with perfect GPS & Compass but noisy actuations and input sensors. Trained till 970 millions, val performance: success: 0.9386(Harsh's 0.97), spl: 0.6536(Harsh's 0.71), softspl: 0.6448 while train performance on train: distance_to_goal: 0.110 reward: 9.105 softspl: 0.774 spl: 0.790 success: 1.000.