Generalizable Task Planning Through Representation Pretraining

The ability to plan for multi-step manipulation tasks in unseen situations is crucial future home robots. But collecting sufficient experience data end-to-end learning often infeasible the real world, as deploying robots many environments can be prohibitively expensive. On other hand, large-scale scene understanding datasets contain diverse and rich semantic geometric information. how leverage such information remains an open problem. In this letter, we propose a learning-to-plan method that generalize new object instances by leveraging object-level representations extracted from synthetic dataset. We evaluate our with suite of challenging inspired household activities (Srivastava, et al. , 2022) show model achieves measurably better success rate than state-of-the-art approaches.