Humanoids 2009
9th IEEE-RAS International Conference on Humanoid Robots

« hide abstract

Object Action Complexes provide a formal framework for the representation of actions and their effects on objects in a cognitive architecture. The framework can be applied on different levels of abstraction, from low-level interaction with physical sensors and effectors to high-level, symbolic planning.
This talk has two parts. In the first part we give a formal definition of Object Action Complexes (OACs). In the second part, we present examples of how this formalism serves to ground object and affordance knowledge in physical experience, and to create symbolic abstractions that can be used for symbolic prediction, reasoning and planning.

« hide abstract

There is a long tradition associating language and other cognitive serial behavior with an underlying motor planning mechanism (Piaget, 1936; Lashley, 1951; Miller, Galanter and Pribram, 1960; Rizzilati and Arbib, 1998). The evidence is evolutionary, neurophysiological, and developmental. It suggests that language is much more closely related to embodied cognition than current linguistic theories of grammar suggest. I'm going to argue that practically every aspect of language reflects this connection transparently The talk discusses this connection in terms of planning as it is viewed in Robotics and AI, with some attention to applicable machine learning techniques.

« hide abstract

This talk discusses work-in-progess in the ROSSI project on building neurocomputational models of affordances and related embodied cognitive mechanisms, and embedding these models in simulated and ultimately physical humanoid robots. This work takes as its starting points the Mirror Neuron System (MNS) II model developed by Michael Arbib and colleagues, which provides a systems-level view of the relevant brain regions and their connectivity, as well as the Chain Model developed by Fabian Chersi and colleagues, according to which neurons encoding subsequent motor acts and leading to a specific action goal are connected in the form of chains of motor primitives. Based on collaborations with neurophysiologists, computational neuroscientists and experimental psychologists in the ROSSI project, current work in our lab involves extending the chain model: (1) to be able to learn motor primitives and adapt to new actions (based on motion capture experiments on action segmentation in humans), (2) to take into account affordances (modeling their processing in the anterior parietal lobe), and (3) to reproduce experimentally observed effects of language processing on motor programs. Finally, the presentation outlines how we plan to integrate these modelling efforts in a (simulated) humanoid robot capable of sensorimotor and social interaction with humans.

« hide abstract

Consider the task of approaching a table in order to grasp an object. A trivial approach to solving this task is choosing a position from which the target of manipulation will be well in reach, and navigating to it. However, a more careful look at the question raises some serious issues: What is a good place in the context of an intended manipulation action? Does the designer's intuition of well-in-reach always imply that the target object can really be reached, given the hardware and control software of the robot?

We address these questions by developing the concept of action-related place, denoted ARPlace. ARPlaces take into account the manipulation and navigation skills of a robot, as well as its hardware configuration. An ARPlace is represented as a probability distribution, that maps the target object's and robot's position to a probability that the target object will be successfully grasped from the corresponding robot position. Therefore, ARPlaces are especially good at dealing with uncertainties that may arise from noisy sensor data. ARPlaces are extremely versatile as they can be merged to allow for multi-joint manipulation, they can be updated as the robot moves to allow for least-commitment planning, or they can be refined to take into account further optimization criteria like time or power consumption.

« hide abstract

We investigate the use of motor invariants, that is, recurring kinematic patterns to improve learning of recognition of actions. Using a multi-subject database of synchronized sensory and motor data in two different domains (grasping and speech) we indentify motor invariants, use them for segmentation and in learning to recognize actions that were possibly never observed before. Experiments show that generalization is improved when motor information is included in the learning machine during training even if it is not available during recognition: i.e. cross-validation is performed without resorting to the original motor data. Further, this advantages are more relevant when the task is more difficult, i.e. because of the addition of noise. Our findings seem to support some of the claims of the motor theory of action recognition (e.g. Liberman, Rizzolatti).

« hide abstract

The concept of object affordances describes the possible ways whereby an agent (either biological or artificial) can act upon an object. The observation of the effects of actions on objects with certain properties allows the agent to acquire internal representation of the world way of functioning, with respect to its own motor and perceptual skills.
Thus, affordances represent the interplay between actions and effects and lie at the core of high level cognitive skills such as planning, recognition, prediction and imitation. Humans learn and exploit object affordances through their entire lifespan, either by autonomous exploration of the world, social interaction or (possibly) introspection.
We propose a computational model capable of encoding object affordances during exploratory learning trials for humanoid robots. We adopt the framework of Bayesian networks and rely on statistical learning and inference methods to generate and explore the network, efficiently dealing with uncertainty, redundancy, and irrelevant information. Interestingly, the model can be applied both to the agents own actions or to those of other indivuduals, in a way similar to the mirror system usually discussed at the sensorimotor level.
Finally, we discuss how this model can be used for different purposes. One such example is imitation learning by exploiting the recognition and planning capabilities to learn new tasks from demonstrations. We show the application of our model in a real world task in which a humanoid robot interacts with objects and uses the acquired knowledge and learns from demonstrations. In addition we will present initial ideas to extend this model to the use of speech and discuss possible future extensions of such an approach.
Related references:

[1] Associating word descriptions to learned manipulation task models, V. Krunic, G. Salvi, A.Bernardino, L. Montesano, J. Santos-Victor, IROS-2008 WORKSHOP on Grasp and Task Learning by Imitation, Nice, France, September 2008

[2] - Learning Object Affordances: From Sensory Motor Maps to Imitation, L. Montesano, M. Lopes, A. Bernardino, J. Santos-Victor, IEEE Transactions on Robotics, Special Issue on Bio-Robotics, Vol 24(1) Feb 2008.

[3] - Visual Learning by Imitation With Motor Representations, M. Lopes and J. Santos-Victor, IEEE Transactions on System Man and Cybernetics - Part B: Cybernetics, June 2005