Also on Google Scholar.
Working paper · October 2026 In progress
Instructions like “set the dining table” name an activity but never state the goal. An LLM infers grounded PDDL goal predicates from a 3D scene graph, the predicates are verified against the planning domain and repaired through a feedback loop, and a symbolic planner produces the plan. The LLM supplies the common sense; the planner keeps the plan correct.
IROS 2026 Workshop on Search Algorithms for Robot Learning · Pittsburgh, PA · September 27, 2026
A sample that is feasible right now can still leave the scene in a state that makes later actions of the same plan harder to refine, triggering failures and replanning. We rank sampled candidates by a learned anticipatory cost and attempt them in increasing order, leaving the sampler, feasibility checks and objective untouched. At a matched sampling budget this cuts planning time by up to 34% and search evaluations by up to 53%.
IEEE Robotics and Automation Letters (RA-L) · 2025
We present a learning-augmented task and motion planning framework that anticipates the future impact of current actions and selects plans that reduce long-term cost across sequences of rearrangement tasks in persistent continuous-space environments.
IEEE International Conference on Robotics and Automation (ICRA) · 2023
We introduce anticipatory task planning, a formulation that helps robots solve current tasks while reasoning about the expected cost of unseen future tasks.
arXiv preprint · 2021
We model floor plans as graphs and use graph neural networks to improve room classification performance by leveraging relational structure in building layouts.