Emre Ugur


From Continuous Manipulative Exploration to Symbolic Planning

This work aims for bottom-up and autonomous development of symbolic planning operators from continuous interaction experience of a manipulator robot that explores the environment using its action repertoire. Development of the symbolic knowledge is achieved in two stages. In the first stage, the robot explores the environment by executing actions on single objects, forms effect and object categories, and gains the ability to predict the object/effect categories from the visual properties of the objects by learning the nonlinear and complex relations among them. In the next stage, with further interactions that involve stacking actions on pairs of objects, the system learns logical high-level rules that return a stacking-effect category given the categories of the involved objects and the discrete relations between them. Finally, these categories and rules are encoded in Planning Domain Definition Language (PDDL), enabling symbolic planning.We realized our method by learning the categories and rules in a physics-based simulator. The learned symbols and operators are verified by generating and executing non-trivial symbolic plans on the real robot in a tower building task. (ICRA2015.pdf) Next, the robot progressively updated the previously learned concepts and rules in order to better deal with novel situations that appear during multi-step action executions. It inferred categories of the novel objects based on previously learned rules, and form new object categories for these novel objects if their interaction characteristics and appearance do not match with the existing categories. Our system further learns probabilistic rules that predict the action effects and the next object states. After learning, the robot was able to build stable towers in real world, exhibiting some interested reasoning capabilities such as stacking larger objects before smaller ones, and predicting that cups remain insertable even with other objects inside. (humanoids2015.pdf)
Robot Videos: ICRA & Humanoids
Symbol Formation

Bootstrapped learning with Emergent Structuring of interdependent affordances

It was shown that human infants develop a set of action primitives such as grasping, hitting and dropping for single objects by 9 months, and later start exploring actions that involve multiple-objects. For example, before 18 months they cannot perceive the correspondence between different shaped blocks and holes and are unsuccessful in inserting blocks in different shapes. This data suggests that there exists a developmental order where infants first develop basic skills that are precursors of combinatory manipulation actions. They also probably use the learned action grounded object properties in further learning of complex actions.
Here we propose a learning system for a developmental robotic system that benefits from bootstrapping, where learned simpler structures (affordances) that encode robot's interaction dynamics with the world are used in learning of complex affordances. We showed that a robot can benefit from a hierarchical structuring, where pre-learned basic affordances are used as inputs to bootstrap the learning performance of complex affordances (ICDL2014-Bootstrapping.pdf). A truly developmental system, on the other hand, should be able to self-discover such a structure, i.e. links from basic to related complex affordances, along with a suitable learning order. In order to discover the developmental order, we use Intrinsic Motivation approach that can guide the robot to explore the actions it should execute in order to maximize the learning progress. During this learning, the robot also discovers the structure by discovering and using the most distinctive object features for predicting affordances. We implemented our method in an online learning setup, and tested it in a real dataset that includes 83 objects and large number of effects created on these objects by three poke and one stack action. The results show that the hierarchical structure and the development order emerged from the learning dynamics that is guided by Intrinsic Motivation mechanisms and distinctive feature selection approach (ICDL2014-EmergentStructuring.pdf).
Robot Video

Brain Mechanisms in Antropomorphic/Non-Antropomorphic Tool Use

A tool (a rake allowing the manipulation of objects out of reach) is grabbed by a trained monkey, parietal neurons with receptive fields on and around the hand expand their receptive fields to represent the tool. Since the human gains the ability to manipulate objects using a robot, can we project from the finding mentioned above that the body schema of a human operator expands to include the robot in the course of controlling it? And if so, what are properties of this body schema expansion. Does the control of all kind of robot morphology induce similar changes in the brain? To investigate these questions we designed an fMRI experiment. Our hypothesis is, that the anthropomorphicity of the robot to be controlled would be critical to how it was represented in the brain.
Tool Use

Reconstruction of Grasp Posture from MEG Brain Activity

We aim at both decoding the grasp type (power vs. precision pinch) and reconstructing the aperture size based on MEG signals obtained while the subjects performed repetitive flexion and extension finger movements corresponding to either precision pinch or power grasp. During the movements, finger joint angles were recorded along with MEG signals using an MEG compatible data glove. For learning the mapping between MEG signals and grasp types, support vector machines with linear kernel is used. For reconstructing the aperture size, a sparse linear regression method is used. (poster)
Grasp Decoding

Learning Object Affordances and Planning in Perceptual Space

The objective of this research topic is to enable an antropomorphic robot to learn object affordances through self-interaction and self-observation, similar to exploring infants of 7-10 months age. In the first step of learning, the robot discovers commonalities in its action-effect experiences by discovering effect categories. Once the effect categories are discovered, in the second step, affordance predictors for each behavior are obtained by learning the mapping from the object features to the effect categories. After learning, the robot can make plans to achieve desired goals, emulate end states of demonstrated actions, monitor the plan execution and take corrective actions using the perceptual structures employed or discovered during learning. (videos)
Affordance Learning

Learning Grasp Affordances through Parental-Scaffolding

Parental scaffolding is an important mechanism utilized by infants during their development. A robot with the basic ability of reaching for an object, closing fingers and lifting its hand lacks knowledge of which parts of the object affords grasping, and in which hand orientation should the object be grasped. During reach and grasp attempts, the movement of the robot hand is modified by the human caregiver’s physical interaction to enable successful grasping. Although the human caregiver does not directly show the graspable regions, the robot should be able to find regions such as handles of the mugs after its action execution was partially guided by the human. (videos)
Parental Scaffolding

Learning Traversability Affordances

We studied how a mobile robot, equipped with a 3D laser scanner, can learn to perceive the traversability affordance and use it to wander in a room filled with spheres, cylinders and boxes. The results showed that after learning, the robot can wander around avoiding contact with non-traversable objects (i.e. boxes, upright cylinders, or lying cylinders in certain orientation), but moving over traversable objects (such as spheres, and lying cylinders in a rollable orientation with respect to the robot) rolling them out of its way. (videos)
Traversability Affordance