Unifying Symbolic Reasoning and Continuous-space Control in Robotics
Abstract
The proposed research project is an ambitious initiative to unify symbolic reasoning and continuous space control in robotics. Despite considerable research toward this objective, existing methods for task and motion planning: (a) use different representations for task planning and motion planning, making it difficult to support bidirectional flow of information and control between the two levels; (b) focus predominantly on kinematics or kinodynamic systems but fail to properly account for the discontinuous dynamics experienced by the robot as it makes and breaks contact with the environment; and/or (c) assume static and known models of planning and control, and limited state uncertainty, making it difficult to adapt to changes in the domain. We posit that to address the limitations of existing work, we need to design a single representation that can be used for both task and motion planning, as well as develop processes that exploit this representation for reliable and efficient reasoning, control, and learning. Our research project will make three key contributions:(i) We will pursue representational unification, with each symbolic relation in an action theory mappedto continuous space in the form of graded concepts # represented with relations and numeric attributes # that hold to differing degrees. Also, each operator will be encoded as axioms that combine symbolic relations with equations for control attributes that consider the dynamics of interactions. These will express functions of the current state#s mismatch with target concepts that are refinedfrom the symbolic level. (ii) We will pursue process unification by interleaving task and motion planning. Symbolic planning will search for a sequence of abstract operators to achieve the desired beliefs with high utility from the current beliefs, considering anabstract action for inclusion in the task plan only if it can find a motion plan, through mental simulations, that accomplishes theabstract action#s outcomes with high utility. During planning and execution, the degree of mismatch between (expected or actual) observations and target concepts will guide adaptation and control. This will also inform both continuous-space and task-space learning.(iii) We will develop an integrated architecture that implements our representational and processing commitments, organizing a setof simple and efficient components such that higher layers constrain lower-level processing and lower layers provide information needed by higher ones. In particular, our architecture will use causal processes to support mental simulations, associate a hybrid variable impedance controller with symbolic skills to handle piecewise-continuous dynamics of robot-environment interactions, and let task planning and motion planning automatically identify and use information relevant to tasks at hand.We will evaluate these methodsthoroughly in both simulation and on physical robot manipulators. To this end, we will focus onscenarios that require a robot to locate and rearrange objects in adesired configuration in the presence of uncertainty, clutter, occlusions, and resource constraints. These will mimic the complexity and context of Navy problems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 15, 2023
- Source ID
- N000142312525
Entities
People
- Mohan Sridharan
Organizations
- Office of Naval Research
- United States Navy
- University of Birmingham