Enactive Robot Learning

Olivier L. Georgeon, Christian Wolf, and Simon Gay 2013. An Enactive Approach to Autonomous Agent and Robot Learning.  IEEE Third Joint  International Conference on Development and Learning and on Epigenetic Robotics (EPIROB2013). Osaka, Japan. August 18-22 2013.

This paper constitutes a short introductory version of our ECA paper. It also presents the experiment of Ernest7 in an e-puck robot. 

Enactive Cognitive Architecture

Olivier L. Georgeon, James B. Marshall, and Riccardo Manzotti 2013. ECA: An enactivist cognitive architecture based on sensorimotor modeling. Biologically Inspired Cognitive Architectures, Volume 6, pp. 46-57, doi: 10.1016/j.bica.2013.05.006. Also presented at BICA2013.

This paper introduces a new way of modeling an agent interacting with an environment called an Enactive Markov Decision Process, inspired by the Theory of Enaction. It also describes Ernest's motivational principle in relation with the autotelic principle (Steels, 2004) and the optimal experience principle (Csikszentmihalyi, 1990). It introduces ECA, the Enactive Cognitive Architecture that drives Ernest 11, and it reports the Ernest 11.2 experiment.

Sensemaking emergence demonstration

Olivier L. Georgeon and James B. Marshall 2013. Demonstrating sensemaking emergence in artificial agents: A method and an example. International Journal of Machine Consciousness, 5(2), pp 131-144, doi: 10.1142/S1793843013500029.

This paper addresses the sensemaking demonstration problem : the problem of demonstrating that an agent gives meaning to or understands its experience. We present a methodology to produce empirical evidence to support or contradict the claim that an agent is capable of a rudimentary form of sensemaking, based on an analysis of the agent's behavior.

As an example, we report an analysis of Ernest's behavior in the Small Loop Problem and we conclude that Ernest is capable of a rudimentary form of sensemaking. This paper is an extended version of our previous paper presented at BICA2012.

Training Ernest 7

This video demonstrates that Ernest develops different behaviors depending on the experience he has during his youth. Here, we have two instances of Ernest: Ernest 1 (brown) is initially kept in the small loop and released on step 290. Ernest 2 (bleu) is confronted to the complex environment right from his birth.

Ernest 1 develops more sophisticated behaviors than Ernest 2 because he is trained to touch both of its sides when he faces a wall. Consequently, after being released, he has a more exploratory behavior than Ernest 2.

Ernest 2 learns to preferably turn to the right when he faces a wall. Consequently, he tends to keep spinning in limited areas of the environment. Ernest 2's learning is limited by the fact that the environment is initially too complex for him to notice sophisticated sequences that involve touching to both sides.

The importance of training is an interesting property of Ernest because it accounts for theories of developmental learning.


(Demo implemented with Ernest r296 and Vacuum r203)

Ernest's source code

Ernest's source code is available here with the instructions to use it. The cleaned-up and tested recommended revision is r296. This revision demonstrates the exact behavior of Ernest 7 reported in the Small Loop Experiment.

Interactional Motivation

 Olivier L. Georgeon, James B. Marshall, and Simon L. Gay 2012. Interactional Motivation in Artificial Systems: Between Extrinsic and Intrinsic Motivation.  In proceedings of the 2nd Joint IEEE International Conference on Development and Learning and on Epigenetic Robotics (EPIROB 2012), San Diego, pp. 1-2.

This paper presents the notion of interactional motivation that drives Ernest, and compares it to reinforcement learning as it is traditionally implemented in Partially Observable Markov Decision Processes (POMDPs).

Ernest 11.5 constructs goals

Like Ernest 11.4, Ernest 11.5 can recognize objects by the possibilities of interaction that they afford. Additionally, Ernest 11.5 has a specific inborn taste for stepping on flowers.

Ernest 11.5 simulates different possible sequences of interactions in spatial memory before selecting the best sequence to enact. These simulations are represented in the bottom-right area of the video. Simulations that produce predictable results (due to information available in spatial memory) are represented with orange outlines. Simulations that produce unpredictable results (due to the lack of information in spatial memory) are represented in blue. The video shows that Ernest learns to simulate increasingly elaborated sequences of interactions as time goes on (see blue squares and triangles spreading in all directions around Ernest from step 253 on).

The high value associated with stepping on flowers favors simulations that lead to even more stepping on flowers. As a result, Ernest learns to make a u-turn to return to a flower when he passes one (see Ernest keeping stepping on the flower from step 260 on).

We find this experiment interesting because it illustrates how an inborn drive can give raise to an explicit goal. Ernest's inborn tendency to step on flowers makes Ernest identify flowers as an interesting goal to reach. Once this goal is recognized, Ernest performs a rudimentary problem-solving computation to reach it. Perhaps the skill to choose a desirable point in space and find a sequence of operations to reach this point underlays higher-level problem-solving skills.

Ernest 11.4 recognizes objects

This video shows Ernest 11.4 learning to interact with different objects in this new version of the Small Loop.

At the beginning, Ernest learns to interact with empty places and with dark-green walls. From step 76 on, he learns to interact with cyan walls. On step 220, we introduce alga, and he starts to learn to interact with them.

Note the funny hesitation on step 234 when Ernest touches an alga for the first time, turns back, and then returns to the alga. Once this new kind of objects is learned, Ernest moves through them without hesitation.

Ernest's previous management of bundles (Ernest 11.2) no longer works in this environment because objects can no longer be identified by disjoint bundles of interactions. Some interactions (e.g., bump) are afforded by different objects (dark-green walls and cyan walls). Ernest 11.4, however, does not actually need to fully recognize objects. He adapts to this environment by only learning "compresences" of pairs of interactions. We borrowed the term compresence from the bundle theory of objects to designate the tie between two interactions that are afforded by the same location in space. In this video, compresences are represented by gray circles that contain interactions (in sequential and spatial memory, top and bottom right areas of the video).

The question of identifying objects by bundles of interactions that are consistently compresent remains an open and difficult question. The notion of compresence seems  to be still controversial in philosophy of objects. Identifying objects raises the question of making analogies between objects, and learning categories of objects based on similarities in the interactions that they afford.

In this experiment:

Touching a cyan wall ahead generates a specific feeling (cyan squares). Touching a cyan wall on the side generates the same feeling as touching a dark-green wall on the side (dark-green squares). Bumping into a cyan wall feels the same as bumping into a dark-green wall (red triangles). Once learned, touching walls ahead "evokes" bumping ahead (light-red triangles in spatial memory, bottom right area of the video). As previously, the evocation of bumping refrains Ernest from trying to move forward towards walls.

Touching an alga ahead generates a specific feeling (light-green squares). Touching an alga on the side generates the same feeling as touching an empty square on the side (white squares). Moving to an alga feels the same as moving to an empty square (white triangles).

(Demo implemented with Ernest r261 and Vacuum r186)