Abstract :
[en] This thesis proposes a new efficient and biologically inspired way of modeling navigation tasks for autonomous mobile robots having restrictions on cost, energy consumption, and computational complexity (such as household and assistant robots). It is based on the recently proposed Reservoir Computing approach for training Recurrent Neural Networks.
Robot Navigation Systems
Autonomous mobile robots must be able to safely and purposefully navigate in complex dynamic environments, preferentially considering a restricted amount of computational power as well as limited energy consumption. In order to turn these robots into commercially viable domestic products with intelligent, abstract computational capabilities, it is also necessary to use inexpensive sensory apparatus such as a few infra-red distance sensors of limited accuracy.
Current state-of-the-art methods for robot localization and navigation require fully equipped robotic platforms usually possessing expensive laser scanners for environment mapping, a considerable amount of computational power, and extensive explicit modeling of the environment and of the task.
This thesis
The research presented in this thesis is a step towards creating intelligent autonomous mobile robots with abstract reasoning capabilities using a limited number of very simple raw noisy sensory signals, such as distance sensors. The basic assumption is that the low-dimensional sensory signal can be projected into a high-dimensional dynamic space where learning and computation is performed by linear methods (such as linear regression), overcoming sensor aliasing problems commonly found in robot navigation tasks. This form of computation is known in the literature as Reservoir Computing (RC), and the Echo State Network is a particular RC model used in this work and characterized by having the high-dimensional space implemented by a discrete analog recurrent neural network with fading memory properties. This thesis proposes a number of Reservoir Computing architectures which can be used in a variety of autonomous navigation tasks, by modeling implicit abstract representations of an environment as well as navigation behaviors which can be sequentially executed in the physical environment or simulated as a plan in deliberative goal-directed tasks.
Navigation attractors
A navigation attractor is a reactive robot behavior defined by a temporal pattern of sensory-motor coupling through the environment space. Under this scheme, a robot tends to follow a trajectory with attractor-like characteristics in space. These navigation attractors are characterized by being robust to noise and unpredictable events and by having inherent collision avoidance skills.
In this work, it is shown that an RC network can model not only one behavior, but multiple navigation behaviors by shifting the operating point of the dynamical reservoir system into different \emph{sub-space attractors} using additional external inputs representing the selected behavior. The sub-space attractors emerge from the coupling existing between the RC network, which controls the autonomous robot, and the environment. All this is achieved under an imitation learning framework which trains the RC network using examples of navigation behaviors generated by a supervisor controller or a human.
Implicit spatial representations
From the stream of sensory input given by distance sensors, it is possible to construct implicit spatial representations of an environment by using Reservoir Computing networks. These networks are trained in a supervised way to predict locations at different levels of abstraction, from continuous-valued robot's pose in the global coordinate's frame, to more abstract locations such as small delimited areas and rooms of a robot environment. The high-dimensional reservoir projects the sensory input into a dynamic system space, whose characteristic fading memory disambiguates the sensory space, solving the sensor aliasing problems where multiple different locations generate similar sensory readings from the robot's perspective.
Hierarchical networks for goal-directed navigation
It is possible to model navigation attractors and implicit spatial representations with the same type of RC network.
By constructing an hierarchical RC architecture which combines the aforementioned modeling skills in two different reservoir modules operating at different timescales, it is possible to achieve complex context-dependent sensory-motor coupling in unknown environments. The general idea is that the network trained to predict the location and orientation of the robot in this architecture can be used to select appropriate navigation attractors according to the current context, by shifting the operating point of the navigation reservoir to a sub-space attractor.
As the robot navigates from one room to the next, a corresponding context switch selects a new reactive navigation behavior. This continuous sequence of context switches and reactive behaviors, when combined with an external input indicating the destination room, leads ultimately to a goal-directed navigation system, purely trained in a supervised way with examples of sensory-motor coupling.
Generative modeling of environment-robot dynamics
RC networks trained to predict the position of the robot from the sensory signals learns forward models of the robot.
By using a generative RC network which predicts not only locations but also sensory nodes, it is possible to use the network in the opposite direction for predicting local environmental sensory perceptions from the robot position as input, thus learning an inverse model.
The implicit map learned by forward models can be made explicit, by running the RC network in reverse: predict the local sensory signals given the location of the robot as input (inverse model).
which are fed back to the reservoir, it is possible to
internally predict future scenarios and behaviors without actually experiencing them in the current environment (a process analogous to dreaming), constituting a planning-like capability which opens new possibilities for deliberative navigation systems.
Unsupervised learning of spatial representations
In order to achieve a higher degree of autonomy in the learning process of RC-based navigation systems which use implicit learned models of the environment for goal-directed navigation, a new architecture is proposed. Instead of using linear regression, an unsupervised learning method which extracts slowly-varying output signals from the reservoir states, called Slow Feature Analysis, is used to generate self-organized spatial representations at the output layer, without the requirement of labeling training data with the desired locations.
It is shown experimentally that the proposed RC-SFA architecture is empowered with an unique combination of short-term memory and non-linear transformations which overcomes the hidden state problem present in robot navigation tasks. In addition, experiments with simulated and real robots indicate that spatial activations generated by the trained network show similarities to the activations of CA1 hippocampal cells of rats (a specific group of neurons in the hippocampus).
