Abstract
Generative models have shown great promise as trajectory planners, giventheir affinity to modeling complex distributions and guidable inferenceprocess. Previous works have successfully applied these in the context ofrobotic manipulation but perform poorly when the required solution does notexist as a complete trajectory within the training set. We identify that thisis a result of being unable to plan via stitching, and subsequently address thearchitectural and dataset choices needed to remedy this. On top of this, wepropose a novel addition to the training and inference procedures to bothstabilize and enhance these capabilities. We demonstrate the efficacy of ourapproach by generating plans with out of distribution boundary conditions andperforming obstacle avoidance on the Franka Panda in simulation and on realhardware. In both of these tasks our method performs significantly better thanthe baselines and is able to avoid obstacles up to four times as large.