Skip to Main content Skip to Navigation

Bootstrapping Robotic Ecological Perception with Exploration and Interactions

Abstract : Robotics has reached a high accuracy on many tasks, like for instance manipulation or navigation. But most of the studies are based on a deep analysis of the problem to solve by the robot designer. These approaches are thus limited to the environment considered by the robot designer, i.e. to a closed environment. Robotics research community is now addressing the issue to allow robots to autonomously achieve tasks in realistic open environments. Such environments are complex and dynamic, like for instance human everyday environment which seems simple but vary a lot from one place to another. In this kind of contexts, the robots must be able to adapt to new situations which were not forecasted by the engineers who designed the robot. Our research work is focused on the development of an adaptive ecological perception for a robotic system. An agent ecological perception defines how it perceives the real world environment through its sensing and acting capabilities. According to J.J. Gibson who has initiated ecological psychology, humans and animals perceive the world through the actions that they can use. Thus, providing a robotic system with the skill to bootstrap autonomously its perception when facing a new unknown situation, would allow the system to be highly adaptive. Our goal is to provide the robot with the capacity to learn a first representation of its surrounding which would work on any environment. This would allow the robot to learn new representations from unknown situations. It is proposed to generate this ability through an interactive perception method. Interactive perception methods take advantage from action to build or enhance representations of the world and then exploit these representations to have more accurate actions. This relation between action and perception can be easily formalized thanks to affordances. Affordance is a concept introduced by J.J. Gibson which is a relationship between visual features, agent skills, and possible effects. The system collects data from an environment by interacting with it thanks to a specific action associated to an expected effect. With these data a probabilistic classifier is trained online to build a perceptual map. This map represents the areas that generate the expected effect when the action is applied. Therefore, the map is called a relevance map. Several relevance maps could be built according to different actions and effects, the sum of these maps represents a rich perception of what the robot can do on its surrounding. We name this final map an affordances map as it allows the robot to perceive the environment through the actions it can use. Our methods was tested on the PR2 robots.
Complete list of metadatas

Cited literature [172 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Thursday, October 15, 2020 - 4:35:33 PM
Last modification on : Saturday, October 17, 2020 - 3:12:59 AM


Version validated by the jury (STAR)


  • HAL Id : tel-02101369, version 2


Léni Le Goff. Bootstrapping Robotic Ecological Perception with Exploration and Interactions. Artificial Intelligence [cs.AI]. Sorbonne Université, 2019. English. ⟨NNT : 2019SORUS219⟩. ⟨tel-02101369v2⟩



Record views


Files downloads