Fig. 4: Diversity results.
From: World and Human Action Models towards gameplay ideation
a, Diversity of three WHAM variants as measured by the Wasserstein distance to human actions. Out of the 102,400 total actions (1,024 trajectories with 100 actions each), we sub-sample 10,000 human and model actions and compute the distance between them. We repeat this ten times and plot the mean ± 1 standard deviation. Closer to the human-to-human baseline is better. Uniform random actions have a distance of 5.3. All models improve through training and can be further improved by up-weighting the action loss. b, Three examples of generations from the 1.6B WHAM produced from the same starting context. We see examples of both behavioural diversity (the player character circling the spawn location versus heading straight towards a Jumppad) and visual diversity (the hoverboard the player character has mounted has different skins).
