WP5. Lifelong improvement of navigation and manipulation skills
The main goal of this package is the design of new algorithms for lifelong improvement of the robot’s capability to coordinately navigate and manipulate objects in its environment. These algorithms will be integrated as a new navigation-manipulation agent that will accept high level goals from the symbolic planner in WP4. The agent will have a limited capacity of planning in its restricted domain of operation. This goal can be divided in the following sub goals and associated tasks:
T5.1. Robust indoor localization with geometric and semantic cues in dynamic environments
Inside well known indoor environments, robust localization is a basic skill that must provide an accurate pose for the robot at anytime. This information is used by most of the other components and agents in the system. However, as in most skills, robust localization in real dynamic environments is only possible if several sources of information are brought together and a final decision is taken according to a given context. A person, or a shifted table should not be confused with a change in the cues used by the localization algorithm. Robust localization in a human habitat heavily depends on the capacity to recognize the objects of the environment and to decide according to what is relevant. This task will build on recent work of UEX, UMA and UCLM to integrate multi-modal odometries with map geometry, inertial corrections and detection of domestic objects and people in a robust probabilistic framework that can improve during extended periods of time.
T5.2 Path planning in a semantically rich space
Our current navigation stack uses Elastic Bands as an adaptable binding between a Probabilistic Roadmap Method and a Rapidly-Exploring Random Tree (PRM+RRT) path planner and a low-level trajectory controller. In this task we will investigate extensions of these schemes to cope with semantically annotated environments. The new planner should be able to generate enriched paths annotated with visual and auditive cues, known objects, recognizable furniture and probabilities of finding humans along the way. Also, a more sophisticated Elastic Band mode will be developed to dynamically project the planned path into the real world as measured by the sensors. Finally, the trajectory controller will also be extended to integrate the head and arms of the robot, so it will be able to follow the path, avoid 3D space obstacles and re-orient, when necessary, to actively perceive the relevant cues.
T5.3. Object picking as an integrated body navigation problem
Arm and hand planning and control are two crucial problems in social robotics. Even without the need of a precise positioning or grasping skill, the reliable picking of everyday objects is a challenging topic in low or middle cost robots. Object picking and placing in narrow environments requires collaborative actions between the robot base and arm to cope with positioning and perception errors. For real time scenarios, planning in the joint space of the base, the arm and the hand is unfeasible. Therefore, we will investigate new lifelong algorithms that will learn the inverse kinematics function of the arm as a free space graph, adapting to the deterioration of calibration conditions and, most importantly, to the appearance of obstacles, such as a table, a human or other objects, avoiding the need of planning safe arm trajectories. These ideas are consistent with goals in T5.3. Here we will pursue a final integration of base, arms and head body parts as a unique coordinated system capable of lifelong learning inverse relations and solve picking and placing problems as a whole.
T5.4 Incremental learning and placing of new objects into a meaningful space representation
Indoor environments suffer frequent modifications mainly due to the actions of humans. Non-static objects like chairs or tables can change their locations, as well as structural elements may modify their aspect. Some environmental modifications are performed in a drastic way, and many others are carried out step by step through small variations. To learn from both types of changes the robot has to reason about the geometry of its environment and the objects in it. This task will investigate new incremental learning algorithms and novelty detection schemes, and integrate them with opportunistic speech queries to the human dweller. As a result, new stable knowledge will be acquired that will have a common grounding with the interacting person. Using these new algorithms, the robot will dynamically learn the relationships between places and objects, establishing important links between locations (e.g. kitchen) and objects (e.g. forks or spoons) even when it has not been initially taught with such information. This task will require tight coordination with WP6 and WP7.
T5.5 Social Navigation and Manipulation
This sub task will explore a topic that is now receiving a lot of attention. It deals with the introduction of sets of constraints to the navigation and manipulation controllers so they become affected by the presence of nearby humans. This task will require tight coordination with WP7 where human activity will be analyzed and fed to the shared internal model, DSR. The derivation and learning of social interaction rules will be facilitated by the rich global state provided by CORTEX. New lifelong learning algorithms will be developed in coordination with WP4 to acquire those rules and to improve human robot interaction.