The goal is the assemble the fuse box. But the required fuse is unreachable to the robot. The robot perceives the turn disc, recalls its past expereince of using it and the ensuing consequnce: the object comes closer, and in such a modified environment the goal is doable. The robot executes a novel plan and realizes the assembly goal.
This task is an approximate recreation of the problem faced by Betty the new caledonian crow when she encountered the task of fetching her dinner basket trapped in a transparent vertical tube (read more). A long wire was available nearby. After a short hesitation, she picked up the wire, fashioned a hook out of it using her beak and used the newly created tool to pull out her dinner basket. Even this little tinge of creativity demonstrated by Betty involved nontrivial “perception-action” and inferring that the goal is directly unrealizable (given the present environment), recall of “a specific” learnt past experience evaluated as valuable in the context of the “present” (making hooks), skill (coupling a potential tool to the body and using it), some basic inference and anticipation of physical causality (i.e. the consequence of pulling the basket up using the newly created “hook tool” connected to the beak) and “flexible” integration of all this knowledge in the context of the goal. We are trying to understand the underleying computational basis of such behaviours!
Darwin Cognitive and Software architecture..
Most of my research in differnet areas like Action, Memory, Learning and Reasoning is integrated together under the umbrella of Darwin Cognitive architecture, presently being used in variety of robots like iCub humanoid, wheeled platforms, Industrial manipulators. The whole architecutre is is composed of multiple modules mainly:
-
Visual Perception (color, shape, 3D localization), Perceptual learning (Objects, Object word association);
-
Action Generation/Forward simulation/Skill learning (PMP framework, Grasping and Grasp Affordances),
-
Object-Action coupling and sensorimotor affordance learning
-
Episodic Memory: Repository of robots past expereinces, enables recall from the present context, future oriented planning and reasoning
-
Observer/Executive Control:Interface to the User of Darwin, Has the working memory, Receives runtime feedback from core modules (Perception-action-memory, body), schedules micro-goals for subsystems, manages conflicts, provides data for the Visualization GUI's of Darwin...
This deliverable of FP7 EU project DARWIN summarizes the integrated cognitive architecture (basically the What, Why, How , Why Now and recent unpublished results). The picture above summarizes the overall organization of the architecture on the iCub platform, that can be configures to other robots (including industrial manipulators as we have done in our work), read more
All the software modules are available open source here, A user manual is found in the documentation section. The Read ME directs the user about where to find what....
This is work in progress :)
Overall, the architecture is domain agnostic (cumulative learning driven, can be deployed for learning multiple tasks and gather further experience), partially embodiment agnostic (can be configured easily to various bodies/ robots).
Contact me, if you indend to test, use further build up the architecture or sub modules on your robot...