Maximally informative interaction learning for scene exploration

Abstract

Abstract-Creating robots that can act autonomously in dynamic, unstructured environments is a major challenge. In such environments, learning to recognize and manipulate novel objects is an important capability. A truly autonomous robot acquires knowledge through interaction with its environment without using heuristics or prior information encoding human domain insights. Static images often provide insufficient information for inferring the relevant properties of the objects in a scene. Hence, a robot needs to explore these objects by interacting with them. However, there may be many exploratory actions possible, and a large portion of these actions may be non-informative. To learn quickly and efficiently, a robot must select actions that are expected to have the most informative outcomes. In the proposed bottom-up approach, the robot achieves this goal by quantifying the expected informativeness of its own actions. We use this approach to segment a scene into its constituent objects as a first step in learning the properties and affordances of objects. Evaluations showed that the proposed information-theoretic approach allows a robot to efficiently infer the composite structure of its environment.

Probabilistic Segmentation and Targeted Exploration of Objects in Cluttered Environments

“…However, if our robot deliberately chooses informative actions, we expect useful object models to be learned faster [35][36][37]46]. Hence, we choose actions to maximize the mutual information I (s; o|a, D T ), where s is the partition of the parts into objects and o is the observed outcome of the action targeted at part a at t = T + 1.…”

Probabilistic interactive segmentation for anthropomorphic robots in cluttered environments

“…Some of these limitations can be addressed by estimating object membership per segment of the image, rather than per pixel [19,21]. Alternatively, the iterative closest point algorithm can be used to find the number of objects that explains the movement in the scene [9].…”

“…However, if the robot deliberately performs pushes it expects to be most informative, segmentations might be obtained faster [19]. Hence, our robot chooses actions to maximize the mutual information I (s; o|a, D), where s is the partition of the parts into objects and o is the observed outcome of an action targeted at part a.…”

“…Movement helps to resolve segmentation ambiguities [13][14][15][16][17]. Rather than waiting for movement in the environment, robots have the possibility to cause changes in the environment themselves [7][8][9][10][11][18][19][20][21], as illustrated in Fig. 1.…”

Extracting kinematic background knowledge from interactions using task-sensitive relational learning

“…they do not address the selection of interaction points for exploration. A promising solution to this problem is to estimate the information gain of actions for minimizing uncertainty has been proposed [10]. Since this kind of methods has not yet fully matured in our context of relational learning, we use a simpler method to generate efficient exploration by rewarding the agent for discovering new knowledge [1], resembling the idea of intrinsic motivation [11].…”