Universal Manipulation Policy Network for Articulated Objects

We introduce the Universal Manipulation Policy Network (UMPNet) – a single image-based policy network that infers closed-loop action sequences for manipulating articulated objects. To infer wide range of trajectories, supports 6DoF representation and varying trajectory length. handle diverse set objects, learns from objects with different articulation structures generalizes to unseen or categories. The is trained self-guided exploration without any human demonstrations, scripted policy, pre-defined goal conditions. support effective multi-step interaction, we novel Arrow-of-Time attribute indicates whether an will change object state back past forward into future. With inference at each interaction step, learned able select actions consistently lead towards away given state, thereby, enabling both goal-conditioned manipulation.