When I am thirsty, I can grab a glas of water without thinking too much about the glas itself. When I run in the woods, I don’t have to closely monitor the ground at every point in time. As I will discuss in more detail in the talk, in both cases, the physical properties of the body (softness & friction of the skin, elasticity in the muscle-tendon system) reduce the computational load for the brain. This is known as Morphological Computation in the context of Embodied Intelligence. In my current research, I am focussing on quantifying the contribution of the physical properties of the body and its interaction with the environment (which are not controlled by the brain) to intelligence. The question is: Is it possible to determine from observations alone, how much of a behavior was actively controlled by the brain and how much of it resulted from physical processes in the body and environment?
I will present information-theoretic approaches to quantifying morphological computation and discuss two applications, namely soft robotics and biomechanics. The talk will close with a presentation of two open theoretical questions that need to be solved to make further progress in this field.
In the last decade Shannon information experienced a revival in the interest of cognitive and Artificial Life modellers. With new conceptual and technical tools available, it has become clear that it offers new opportunities to analyse, predict and construct cognitive models in a principled way, while reducing assumptions about the particular cognitive architecture of interest.
Several of these have already received wide interest, such as rate/distortion-type arguments ("relevant information") for decision-making regularization, or predictive information and impowerment as intrinsic motivation mechanisms.
In my talk, I would like to argue that information theory offers much more than just being a specialized tool to formulate specific models for cognitive dynamics: in fact, its use means adopting a whole philosophy about modelling cognition and its development in nature. The talk will present a collection of examples (some more, some less known) for that point, including a sensor evolution scenario that sheds new light on a non-obvious observation from biological evolution.
Intelligent agents, interacting with their environment, operate under constraints on what they can observe and how they can act. Unbounded agents can use standard reinforcement learning techniques to optimize their inference and control under purely external constraints. Bounded agents, on the other hand, are subject to internal constraints as well. This only allows them to partially attend to their observations, and to partially intend their actions, requiring boundedly-rational selection of perception and action.
This problem is particularly interesting when restricted to memoryless agents. Their optimization is a sequential rate-distortion problem of trading off internal communication costs with external costs. The solution exhibits intriguing and useful phenomenology, such as phase transitions which cluster observations into actions in discrete domains, or reduce the controller order in continuous domains. Moreover, the general problem with retentive (memory-utilizing) agents can be reduced to the memoryless case by considering communication costs on the memory channel.
Human and animal movement is absolutely fascinating and can hardly be mimicked by technical devices, so far. It has been proposed that part of the movement generation and control can be attributed to the non-linear characteristics of the bio-mechanical structures and the morphology. Terms like morphological computation (Paul, 2006) and intelligence by mechanics (Blickhan et al., 2007) capture this idea. With the aid of computer simulations of human movement, we investigate the contribution of muscles and other structures to the movement generation.
The talk will present the simulation approach (multi-body Lagrange, muscle models, and neural motor control concept) and discuss methodical overlap with non-linear dynamics and information theory for the evaluation of the model results. Finally, some applications in medical engineering and automotive ergonomics will be presented.
Human and animal movement is absolutely fascinating and can hardly be mimicked by technical devices, so far. It has been proposed that part of the movement generation and control can be attributed to the non-linear characteristics of the bio-mechanical structures and the morphology. Terms like morphological computation (Paul, 2006) and intelligence by mechanics (Blickhan et al., 2007) capture this idea. With the aid of computer simulations of human movement, we investigate the contribution of muscles and other structures to the movement generation.
The talk will present the simulation approach (multi-body Lagrange, muscle models, and neural motor control concept) and discuss methodical overlap with non-linear dynamics and information theory for the evaluation of the model results. Finally, some applications in medical engineering and automotive ergonomics will be presented.
Many biological microswimmers, including sperm and motile green alga, navigate in noisy environments. For successful navigation, these swimmers integrate their active motion and sensory perception in tight feedback loops, relying on only minimal computational resources. We first discuss a geometric sampling strategy along helical paths that allows sperm cells from marine invertebrates to find the egg, even when the read-out of chemical orientation signals is extremely noisy. Swimming along curved paths allows these cells to actively probe their environment in a stereotypic manner, thereby structuring the type of sensory information the cells will perceive.
As a second example of intelligent motility control, we will present a purely mechanical feedback mechanism that allows a breast-swimmer-alga to coordinate its two flagellar swimming arms. Physical interaction with the surrounding fluid is essential to ensure a synchronized swimming gait. The robustness of these biological control mechanisms is reviewed in the presence of active fluctuations and extrinsic noise, which can be significant at the scale of cells.
There is a significant performance gap between human agents and robotic agents. Closing this gap has been the goal or robotics research for many years now. In this talk, I would like to speculate on desirable properties of solutions capable of closing that gap (as opposed to the development of increasingly competent skills that serve specific applications but do not achieve the generality required to be part of closing the gap). I will propose three such desirable properties and support my view with experiments from the areas of grasping, interactive perception, and learning from (inter-)action. As these experiments can only be considered circumstantial evidence for my claims, I prefer to refer to these properties as trips and tricks.
