monash_171915.pdf (3.87 MB)
Learning manipulation affordances via intrinsically motivated exploration
thesisposted on 2017-03-02, 01:56 authored by Border, Benjamin James
In this thesis, the problem of unsupervised discovery and learning of manipulations for an object is addressed. The concept of manipulation affordances is introduced, which characterise a naive model of an object's expected motion due to interaction, given various conditions. Interaction with an object is performed via pushing, sliding or dragging motions. In order to learn the motion of an object, the agent observes how the object itself moves as a result of the agent's own interactions with it, then generalises and learns the necessary characteristics and conditions to recreate the object motion. Generalisation requires knowledge of the environment the object resides in. This thesis describes the environment via constraints, which represent impenetrable barriers for the object. Generalisation is then performed by referencing affordances from constraints, recognising symmetry and parametrisation. It is also necessary to be able to use affordances to complete tasks. This is addressed by generating a discrete sequence of moves from affordances using a recursive tree algorithm. It is shown that affordances of an object can be generalised and learned, then used to perform simple manipulation tasks for an object. Discovery plays an important role in building a repertoire of manipulation affordances. The ability to discover and learn new affordances can be useful for unfamiliar environments or objects and enables a measure of adaptability. Verifying and revisiting discovered, but unfamiliar affordances, can also benefit the robotic agent by refining observed characteristics. Random exploration or a specified strategy can be used for discovery and refining of affordances, however these methods can be sub-optimal, prone to unnecessary repetitions or suffer from poor adaptability. Another approach, is to use intrinsic motivation in the form of interest to direct the robotic agent's actions. In this thesis, the agent uses intrinsic motivation to prioritise interaction in areas of interest, where interest is generated from the novelty of known affordances of the object. It is shown that intrinsic motivation can be used to facilitate exploration of affordances, that it is adaptable and improves over random exploration. This thesis experimentally evaluates the process of affordance discovery, learning and use in tasks. Evaluation is conducted in 2D environments via a real robotic system consisting of a fixed 3 degrees of freedom finger with a compliant 1D tactile sensor and a vision system for tracking the object. Simulations are also conducted, where some limitations of the fixed 3 DOF finger are addressed. The performance of motivated exploration approaches in this thesis to that of random exploration is also experimentally compared.