Reaching Through Latent Space: From Joint Statistics to Path Planning in Manipulation

We present a novelapproach to path planning for robotic manipulators, in which paths are produced via iterative optimisation the latent space of generative model robot poses. Constraints incorporated through use constraint satisfaction classifiers operating on same space. Optimisation leverages gradients our learned models that provide simple way combine goal reaching objectives with satisfaction, even presence otherwise non-differentiable constraints. Our trained task-agnostic manner randomly sampled In baseline comparisons against number widely used planners, we achieve commensurate performance terms task success, time and length, performing successful obstacle avoidance real 7-DoF arm.