Learning Better Paths: Multimodal Generative Models Enhanced with Local Critics

This work proposes a novel framework to accelerate motion planning in previously unseen environments with obstacles by modeling the distribution of the collision-free configuration space using Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP). To effectively manage multimodal data, we condition the WGAN-GP using a Variational Autoencoder (VAE) embedded in a continuous latent space. The configuration space is approximated through a set of Gaussian distributions, allowing the dataset to be segmented into multiple localized models. This strategy not only enhances the learning efficiency but also reduces convergence time. We utilize the reconstructed configuration space to evaluate motion planning performance in previously unseen scenarios. Experimental results highlight the potential of our approach to significantly accelerate planning in unknown environments while maintaining the generation of near-optimal trajectories.