Learning, Adapting & Evolving

evolving

Organisms and natural systems have not only evolved numerous functions (such as vision and self-healing) but have acquire the ability to learn, adapt and evolve. This enables them to stay functional and operate optimally despite the changes in the environment. This section groups EPFL’s research projects that take technology to a new level by making it adaptable.

Biomimetic behavior of an active tensegrity structure

Biomimetic structures interact with their environment, change their properties, learn and self-repair, thereby providing properties that are similar to living organisms. Interactions with the environment involve unique challenges in the field of computational control, algorithms, damage tolerance, and structural analysis. Tensegrity  structures are pin-jointed structures of cables and struts in a self-stress state. Tensegrity structures are suitable for active control since the shape of the structure can be changed by changing the length of the elements. When a tensegrity structure integrates sensors, actuators (devices that allow structural members to change their length) and a control system, deployment and modification of structural behavior becomes feasible. We are developing active control systems for tensegrity strucutres that are capable of self-diagnosis, self-adaptation and learning.

N. Veuve, S.D. Safaei and I.F.C. Smith, Towards development of a biomimetic tensegrity footbridge

Research Lab: prof. Ian Smith, Applied Computing and Mechanics Laboratory

 

Neural mechanisms of movement and learning for novel robots capable of agile locomotion in complex environments

We work on the computational aspects of movement control, sensorimotor coordination, and learning in animals and in robots. We are interested in using robots and numerical simulation to study the neural mechanisms underlying movement control and learning in animals, and in return to take inspiration from animals to design new control methods for robotics as well as novel robots capable of agile locomotion in complex environments.

Research Lab: prof. Auke Ijspeert, Biorobotics Laboratory

 

Bio-inspired and Neuromorphic Circuits

The operation of the central nervous system offers a very contrastive example to classical processor architectures’. Consuming approximately 20W, the brain accomplishes task that are yet accepted very difficult to classical processors, such as complex and multimodal pattern memorization and classification, complex decision making and consciousness. Classical processors architectures appear to lend themselves to solving simple arithmetic or logic sequential tasks, in an aggressively repetitive approach. Novel bio-inspired computational methods may take benefit of the silicon media implementation to carry out complex tasks in a power-efficient manner. The intrinsic characteristics of current or future devices may be used, circuits and systems developed, and finally processor architectures may be optimized to support bio-inspired algorithms.

On the other side, electronic systems have taken benefit of the stable CMOS fabrication technology, as well as suitable processor architectures to expand into ubiquitous industrial, medical and consumer products. In spite of an evident mature technology, the microelectronic industry must prepare to face stringent limits related to the power dissipation for a given operation, with strong impact on the circuit size, the operation speed, the global required energy, and eventually the marketability of new products.

The research and projects cover areas from the usage of new devices to emulating bio-inspired behavior, microelectronic circuits and systems realizing the hardware implementation of nature-inspired computational principles (e.g., early image processing at the sensor level, stochastic resonance, unconventional computing), architectures for artificial neural networks and deep learning, and reliability of neural operators realized in microelectronic hardware.

Research lab: dr. Alexandre Schmid, Microelectronic Systems Laboratory

 

Evolutionary Robotics – RobGen

The promise of Evolutionary Robotics to completely automatize the design of robot controllers and/or morphologies is an idea with great appeal not only to researchers, but also to students. Recently, we introduced the RoboGenTM an open-source software and hardware platform for Evolutionary Robotics, and described its success as an educational tool in a masters level course at EPFL. There it was shown that RoboGen could provide students with valuable hands on experience with Evolutionary Robotics, neural networks, physical simulation, 3D printing, mechanical assembly, and embedded processing.

Research Lab: prof. Dario Floreano, Laboratory of Intelligent Systems

 

Robogami – next paradigm of robots with augmented adaptability to conform to the unexpected and ever changing environment

Robotic Origamis (Robogamis) are functional robots constructed from smart materials that embed actuation, computation, communication, sensing, and interface with other components such as specialized sensors. Robogamis are designed and fabricated as flat sheets that can adapt to their environment by transforming their form using embedded actuators, sensors and controller.

Research Lab: prof. Jamie Paik, Reconfigurable Robotics Lab

 

Evolution and Computation

Evolution can be viewed as a form of computation. This viewpoint allows us to study questions related to the nature of the steady state of evolution, its robustness and efficiency using tools and techniques from computer science. This allows us to further our understanding of evolutionary mechanisms that enable organisms fast adaptations and have generated the diversity of life we see today.

References:

Ioannis Panageas, Piyush Srivastava, Nisheeth K. Vishnoi, 2016., Evolutionary dynamics on finite populations mix rapidly. (PDF)

Nisheeth K. Vishnoi, 2015., The speed of evolution. (PDF)

Nisheeth K. Vishnoi , 2013. Making evolution rigorous- the error threshold. (PDF)

Narendra M. Dixit, Piyush Srivastava, Nisheeth K. Vishnoi, 2012., A finite population model of molecular evolution: Theory and computation. (PDF)

Kushal Tripathi, Rajesh Balagam, Nisheeth K. Vishnoi, Narendra M. Dixit, 2012., Stochastic Simulations suggest that HIV-1 survives close to its error threshold. (PDF)

Research lab: prof. Nisheeth Vishnoi, Theory of Computation Laboratory 3