Learning task outcome prediction for robot control from interactive environments

Robots, Pancakes, and Computer Games

“…It is hard to quantify what makes "good" motion data for PbD. Based on the experience gained in previous work [5,11,16], three aspects of motion data were identified to operationalize data quality for the purposes of assessing H2.…”

“…In the work of Haidu et al [11] the authors learn a failure detector model to allow a robot to recognize the point where the current action will lead to a failure. The data used for learning is collected from a virtual environment by running multiple episodes (success and failure cases) of the given action.…”

“…Playful tools for motion data acquisition based on simulated virtual environments have been explored in the robotics community [11,16]. They offer the benefits of a fully observable world and safety from real world constraints.…”

“…It includes complex and diverse manual activities that are easy to perform for (many) people but are hard to describe in a formal or algorithmic way. In robotics, cooking simulators are being investigated for knowledge acquisition [3,5,11,16]. We chose cooking as a paradigmatic scenario for an activity in the household and demonstrate how human computation games that go beyond the playfulness of simulator tools can help to gather data from ordinary persons who have no experience in robotics.…”

“…We thus need to map task domain actions to game domain intentions (Table 1). Since cooking is an activity for which a vast amount of behavior can be identified, we chose pancake making as a well-defined example also used in related research [3,11,16]. We assume a simplified procedure that uses ready-made pancake batter, reducing the task to few simple actions ( Table 2).…”