A Formal Account of Structuring Motor Actions With Sensory Prediction for a Naive Agent

Abstract

For naive robots to become truly autonomous, they need a means of developing their perceptive capabilities instead of relying on hand crafted models. The sensorimotor contingency theory asserts that such a way resides in learning invariants of the sensorimotor flow. We propose a formal framework inspired by this theory for the description of sensorimotor experiences of a naive agent, extending previous related works. We then use said formalism to conduct a theoretical study where we isolate sufficient conditions for the determination of a sensory prediction function. Furthermore, we also show that algebraic structure found in this prediction can be taken as a proxy for structure on the motor displacements, allowing for the discovery of the combinatorial structure of said displacements. Both these claims are further illustrated in simulations where a toy naive agent determines the sensory predictions of its spatial displacements from its uninterpreted sensory flow, which it then uses to infer the combinatorics of said displacements.

From State Transitions to Sensory Regularity: Structuring Uninterpreted Sensory Signals From Naive Sensorimotor Experiences

“…Such a (slight) generalization allows to highlight how a specific set of motor trajectories actually condition the sensory transitions available in the agent's sensorimotor flow. More precisely, we introduced in [16] the sensor receptive field as the specific region of space for which the state of the environment is sufficient to fully determine the agent's sensory state s. Formally, a sensor receptive field can be seen as a function…”

“…noise, imperfect repetition of motor paths, etc.? The very same limitation is also shared by the previous work by the authors [16]; in this contribution, the interlink between motor actions and sensory prediction is explored, through the demonstration of the existence of a group isomorphism between them. But predicting the sensory outcome of an action is only accessible to the agent by detecting the exact shift of values inside its own sensor array.…”

“…sensory S) set. As shown in [16], the agent's motor 174 description can be enriched from m ∈ M to b ∈ B, where 175 b = (m, τ ) depicts the absolute agent's motor configuration. 176 b is made of the agent's internal (and thus known to it) 177 motor configuration m and of its absolute external (and thus 178 unknown to it) pose τ in its ambient space.…”

“…176 b is made of the agent's internal (and thus known to it) 177 motor configuration m and of its absolute external (and thus 178 unknown to it) pose τ in its ambient space. Importantly, 179 as discussed in [16], switching the motor description from 180 m to b allows us to keep a functional relationship between 181 motor and sensory data, even in the case where the agent 182 can freely move in its environment. But while the agent has 183 no direct access to b, it can apply some motor actions a on 184 b = (m, τ ) to go to configurations b ′ = (m ′ , τ ′ ) = ab: the 185 agent knowns instead how to move in B.…”

“…More precisely, the spatial and temporal dynamics of the agent's exploration is shown critical to obtain a meaningful and useful structure of its own sensory symbols. Next, some experiments initially proposed in [16] are reproduced in Section V to illustrate how the proposed framework could allow an agent to actually build some sensory prediction functions even in the presence of 152 sensory noise. Finally, a conclusion ends the paper.…”