Adaptation of Task Goal States from Prior Knowledge

TL;DR

This work addresses the challenge that task goal states in robotics are often not fixed but admit variation. It introduces a formal framework based on ValueDomains and Variations to represent goal states as subsets of environmental properties, derived from a single demonstration through interactive disambiguation. The method defines Actions and Skills as the abstract and concrete means to modify environment properties, and outlines a five-step planning-execution pipeline that computes differences, generates executable plans, and dispatches skills to agents. The approach is demonstrated in simulation with a Panda robot, showing how goal variations can be discovered and used to plan feasible sequences of actions; limitations related to multiple-instance variations and precondition handling are discussed, with clear paths for future improvement and extension.

Abstract

This paper presents a framework to define a task with freedom and variability in its goal state. A robot could use this to observe the execution of a task and target a different goal from the observed one; a goal that is still compatible with the task description but would be easier for the robot to execute. We define the model of an environment state and an environment variation, and present experiments on how to interactively create the variation from a single task demonstration and how to use this variation to create an execution plan for bringing any environment into the goal state.

Figures (9)

• Figure 1: Understanding, i.e. representing, the intended amount and type of variation in a task's goal state is paramount for task monitoring and execution planning.
• Figure 2: Variations can be used to express a desired range of values for environment states. On the left are the contents of an environment state. The right part shows RangeVariations of the ValueDomains on the left. $A \rightarrow B$ means $A$ is a subtype/-concept of $B$. Bold types are variations.
• Figure 3: The Container concept defines the contentLevel property as a Number. This property has a value of $0.45$ in the WhiteMugInstance. One (or more) Skill(s) must be executed to bring the current level to the desired level inside the defined variation on the right.
• Figure 4: From a single user demonstration, the system extracts the desired task goal state with the help of user interaction to solve ambiguities. Using the created environment variation, the system computes a task execution plan to bring new environments into the goal state. It sends the plan to agents in the environment to execute.
• Figure 5: The goal state is a RangeVariation of the environment, of type EnvironmentDataRangeEntityVariation, which contains a variation of entities. This sub-variation is a RangeVariation of type MapRangeInstanceSubset (variation of type $A$, see \ref{['ssec:variations']}) and contains one instance RangeVariation of type InstanceRangePropertiesVariation. It defines the instance's concept RangeVariation, a LiquidContainer to be found in the environment, and the contentLevel property RangeVariation, the closed interval $\left[0.28, 0.32\right]$.