Commonly used tools, hand written scripts, works of art and architecture are products of skilled human movement. Given the staggering diversity of human actions and ensuing creations, it is intriguing to investigate if there are common repeating patterns, primitives or building blocks that underlie their construction. The line of research explores this issue in the context of an integrated action ‘perception-generation-simulation’ framework being developed for the iCub humanoid, with specific focus on a) fast,
Learning Diverse motor skills: Fast, Green and Cumulatively
The video shows a more recent experiment on iCub learning to use a turn disc tool (inspired from similar animal reasoning tasks), by spontaneously imitating the teacher. Further the conseqence of such tool use (i.e the fuse i.e the white obejct comes closer or becomes reachable) is also abstracted and encoded in the memory.
This experiemnt shows how the robot recalls it past experience of the affordance of the turn disc in the context of an otherwise unrealizable goal: Assembly of a Fuse box. The task itself is an approximate recration of the creative behaviour demonstrated by the caledonian crow Betty when her food basket was trapped in a vertical tube. The analogy here is Betty-Robot, Food Basket-Assembly Goal (that is impossible as one of the component needed is unreachable), piece of wire-Turn Disc. The robot infers the goal is unrealizable, recalls past experience in the context of the present environment (and simulates the ensuing conesqunce), generates a novel behaviour: Grasp the handle-use the tun dis-track the fuse- grasp the fuse-insert it in the fuse box.
cumulative learning of different motor skills, tool use actions by the robot; b)Combining multiple streams of learning: imitation, physical interaction, motor imagery; c) facilitating task-independent motor knowledge abstraction/reuse during such learning and; d) task-specific compositionality during new tasks, contexts.
A central, novel feature of the computational framework is that both the action-perception and action-generation/imitation loops operate at the level of movement ‘shape’ (in contrast to movement trajectory). The rationale is to liberate the movement trajectory from task-specific details (of who i.e. end effector/tool; where i.e. spatio-temporal constraints; when; why i.e. ensuing consequences), and store learnt motor knowledge in an abstract fashion, hence offering the learner (i.e. humanoid here) with the implicit potential to spontaneously ‘perceive, imitate and begin to perform’ several other skills (that share a similar structure).
With different experiments and results both on humans and humanoid robot, our work shows that a range of common tool-use actions, hand-written scripts (Devanagari, Latin etc.), and architectural designs can be learnt and generated by a complex redundant body by combining only three basic shape features: line, bump and cusp. Building up, the iCub skill learning architecture also proposes, a principled approach to learn, represent and organize (in a distributed fashion) different aspects of motor knowledge like a) the desired spatio-temporal motion of the end effector/tool during an intended skill, tool use action; b) geometric/causal relation between the body, coordinated tools (and other bodies) c) attaining specific postures during diverse tasks, while facilitating exploitation of such information in novel goals, learning scenarios.
Going the other way round (humanoid to human) and connecting topics like the apparent user-friendliness of a potential novel tool itself, the ability to spontaneously imitate (with a fair enough first prototype), potential ingredients of procedural memory in human (and humanoids), emerging trends from motor neurosciences related to action observation-simulation-generation networks in the brain in the context of the proposed framework, we ask whether human skills, tools, scripts and artistic creations are the way they are because the brains language of movement is spoken through the vocabulary of ‘shapes’.
Mohan, V., Morasso, P., Zenzeri, J., Metta, G., Chakravarthy, V,S., Sandini, G. (2011). Teaching a humanoid robot to draw 'Shapes'. Autonomous Robots, 31(1), pp. 21-53.
Mohan V and Morasso P (2011) Passive motion paradigm: an alternative to optimal control. Front. Neurorobot. 5:4. doi: 10.3389/fnbot.2011.00004.
Mohan V, Morasso P (2012) How past experience, imitation and practice can be combined to swiftly learn to use novel “tools”: Insights from skill learning experiments with baby humanoids. Intl. Conf on Biomimetic and Biohybrid systems: Living Machines 2012, July 9-12 2012, Barcelona,Spain.
Morasso P, Casadio M, Mohan V, Rea F, Zenzeri J (2015) Revisiting the body-schema concept in the context of Whole-Body Postural-Focal Dynamics. Frontiers in Human Neuroscience, doi: 10.3389/fnhum.2015.00083.
Orban, G. A., & Caruana, F. (2014). The neural basis of human tool use. Frontiers in Psychology, 5, 310. doi:10.3389/fpsyg.2014.00310
Mohan V, Morasso P, Sandini G (2014). Towards Brain guided cognitive architectures. Book Chapter in 'Bioinspired approaches for human centric technologies'. Editor R. Cingolani, Springer International Publishing Switzerland, DOI 10.1007/978-3-319-04924-3_7.
Cotugngo G, Mohan V, Nanayakkara, T, Althofer K. (2014). Simplifying Grasping Complexity through Generalization of Kinaesthetically Learned Synergies. IEEE-ICRA 2014, Hong Kong.
Diedrichsen, J., & Kornysheva, K. (2015). Motor skill learning between selection and execution. Trends Cogn Sci.