Temel Kavramlar
This thesis presents novel algorithms that efficiently solve complex Task and Motion Planning (TAMP) problems by tightly integrating discrete task planning with continuous trajectory optimization, adaptively combining sampling and optimization methods, and accelerating computations using deep learning.
Özet
The thesis focuses on improving the performance of Task and Motion Planning (TAMP) algorithms from three complementary perspectives:
-
Integrated Planning and Optimization for Task and Motion Planning:
- The author investigates the integration of discrete task planning with continuous trajectory optimization.
- The main contribution is a conflict-based solver that automatically discovers why a task plan might fail when considering the physical constraints, and feeds this information back into the task planner.
- This results in an efficient, bidirectional, and intuitive interface between task and motion planning, capable of solving TAMP problems with multiple objects, robots, and tight physical constraints.
-
Meta-Solvers: Adaptive Combination of Sampling and Optimization Methods:
- The author illustrates that neither sampling nor optimization is superior in all TAMP settings.
- To combine the strengths of both approaches, meta-solvers are designed - adaptive solvers that automatically select which algorithms and computations to use and how to best decompose each problem.
- This allows finding solutions faster by adapting the solver to the specific problem at hand.
-
Accelerated Task and Motion Planning with Learning Methods:
- The author investigates how to use deep learning to accelerate computationally expensive operations within TAMP solvers.
- Specifically, deep generative models are proposed to provide good initial solutions for nonlinear optimization, and graph neural networks are used to directly predict which constraints are infeasible.
- A key contribution is leveraging the factored structure of TAMP problems to create more accurate and generalizable neural models.
Throughout the thesis, a refined, factored representation of the trajectory optimization problems within TAMP is used, which exposes the local dependencies and enables more efficient planning, encoding of geometric infeasibility, and meta-reasoning.