Learning Ability Models for Human-Robot Collaboration

Our vision is a pro-active robot that assists elderly or disabled people in everyday activities. Such a robot needs knowledge in the form of prediction models about a person’s abilities, preferences and expectations in order to decide on the best way to assist. We are interested in learning such models from observation. We report on a first approach to learn ability models for manipulation tasks and identify some general challenges for the acquisition of human models. I. PLAN-BASED CONTROL FOR HUMAN-ROBOT COLLABORATION We are interested in planning and plan execution mechanisms that allow a robot to actively participate in joint tasks with a human partner, especially in assistive scenarios. We assume that in such a task there will only be limited explicit communication similar to the way when humans participate in a shared task. One crucial factor in the development of such joint planning abilities is knowledge about the capabilities and preferences of the partners involved. When helping a person, a robot should take over those tasks that are difficult to perform for the person, but feasible for the robot. For example, an elderly person with a walking impairment can be supported by bringing items she needs to prepare a meal. But the robot should not attempt to cut the ingredients: first, current robots are not capable of doing such sophisticated manipulation tasks reliably, so the probability of failing would be very high and second, the robot should leave work to do for the person to keep her active and healthy and not to intrude into her private life more than necessary. There is a wide variety of models that are interesting for a robot to interact with a human. Models describing the workspace of robots and humans [4] can be used to coordinate actions in joint workspaces or to choose high-level actions that avoid spatial conflicts. Another approach is to model criteria of social comfort into planning algorithms [1]. In our work, we try to describe capabilities on an action level, for example if a specific manipulation task would succeed in a given situation. We are interested in criteria such as success probability, effort for a person, efficiency and social acceptability (for example it might not be appropriate for a robot to touch food). Because capabilities and preferences vary strongly among individuals, we would like a robot to learn models about itself and its human partner from experience. In the following we present our observations from a first attempt to learn the capabilities of agents to pick up and put down objects. We report on the challenges that we found in this work and summarize them in a general way to identify problems for similar learning problems. II. LEARNING ABILITY MODELS For our research on joint human-robot activities we use a physical simulation of two agents (that are both displayed as robots): one is acting as an autonomous robot, the other one is controlled by a human via the keyboard. A person can move such an agent freely in the world and give commands for gripping and putting down objects. The gripping and put down actions are executed autonomously based on heuristics. These manipulation actions are implemented in different ways for the autonomous robot and the human-controlled one, so that the capabilities are not identical. In a user study [2] we acquired data from nine subjects who had the task to set and clear the table in two simulated kitchen environments. In total, we observed about 60 gripping and put-down tasks respectively from the execution of complete plans for each participant. For learning capability models, we assumed that all participants were equally skilled in the manipulation tasks, because those were executed autonomously. This is not completely true, because the success of the task also depends on the position where the agent is standing while gripping. But we doubted that any learning algorithm would succeed with the small number of samples we had for each participant and with this simplification we had around 700 examples in total (This includes gripping and put-down tasks that were performed in incomplete runs that were aborted for some reason. This data was not used for evaluation in the user study, but can be used for our purposes here). Beside the data from the user study, we also collected analogous data for an autonomous robot. As a first approach, we tried to learn prediction models of when such tasks succeed or fail by using decision trees (using the Weka J48 algorithm), for example this function for putting down an object: object-goal-position× object-type→ success/failure. The object positions were given relative to the piece of furniture they were standing on or had to be put on, which was general enough given the samples from predefined scenarios from our user study. The result of these learning attempts was not surprising, but still disappointing: the decision tree judged that both the robot and the humancontrolled agent will succeed in all cases. Looking at the learning experience, the reason for this result was very obvious: only about 4% of the gripping tasks from the user study were identified as failures and 7% of put-down task. The rates for the autonomous robot are similar. Beside the low failure rate overall, there was no obvious structure when manipulation tasks succeed and fail. One hypothesis was that the wall behind the worktop would be