Behavior Composition

Visualization of the measure values resulting from a sequence of four randomly selected desired measures (left) and text labels (right). The desired behavior in both cases is changed four times throughout the episode. The measure values from a policy sequence (Section 4 of the paper) are shown on the left as a function of time. These were run 10 times and the corresponding error plots are shown. For both measure and language conditioning, we found that our generated policies and their resulting behavior sequence closely match the desired behavior sequence. We found that our model is capable of generating and sequencing diverse policies despite large changes in foot-contact time and sudden changes in behavior between trajectory segments. By labeling policies that fall over as such, our language-conditioned diffusion model is capable of pruning away poorly-performing policies and dramatically increasing our performance in sequential behavior composition.