Embedded machine learning

We have started developing sensor systems to explore the interface of computationally-constrained microcontrollers and high-performance cloud computing. In a first example, we collected training samples (of collisions, movement, and rotation) from our sensor system and subsequently trained a time series classifier on a conventional computer. After training, we implemented the algorithm on a low-power microcontroller (SAMD21E), and demonstrated that it could successfully discriminate between the various events. We intend to expand this work and augment our air-releasable robots (ARRs) with high functionality at low-cost by taking advantage of modern machine learning including deep learning techniques. By developing our own robotic systems, we control the type of data we generate and possess a unique intuition for the training process.

  • Multi-Functional Sensing for Swarm Robots Using Time Sequence Classification: HoverBot, an Example
    M.P. Nemitz, R.J. Marcotte, M.E. Sayed, G. Ferrer, A.O. Hero, E. Olson, A.A. Stokes
    Frontiers in Robotics and AI, 5, 55

Using inexpensive microcontrollers for multi-functional sensing. We started to implement signal processing techniques that have been widely used for voice recognition in criminal cases, to discriminate distinct time series from one another. In this case, a low-cost microcontroller and a trained algorithm were capable of discriminating between collision, rotation, and movement.

Embedding electronics in soft materials

We have developed robotic systems that demonstrate the combination of electronics and soft materials. We developed Wormbot, a soft robot that uses voice coils for actuation (locomotion) and communication. We particularly focused on fully-integrated systems in which all components resided within the soft body, as opposed to systems in which control elements are external to the robot.

  • Using Voice Coils to Actuate Modular Soft Robots: Wormbot, an Example
    M.P. Nemitz, P. Mihaylov, T.W. Barraclough, D. Ross, A.A. Stokes
    Soft Robotics, 3(4), 198-204
  • Linbots: Soft Modular Robots Utilizing Voice Coils
    R.M. McKenzie, M.E. Sayed, M.P. Nemitz, B.W. Flynn, A.A. Stokes
    Soft Robotics, 6(2), 195-94

Sequential actuation of electro-magnetic soft actuators. A series of frames showing one cycle of sequential expansion and attraction of

We can also effectively use the combination of electronics and soft materials for the development of sensor systems. In one example, we integrated soft channels (which we subsequently filled with liquid metal) and a RFID chip into an elastomeric polymer. This device is a wireless stretch sensor in which the soft channels act as antennas. Stretch leads to a change in resistance (of the antenna), which leads to a shift in resonance frequency of the RFID circuit. Our system does not require external power, which makes it extremely suitable for the development of inexpensive sensors.

  • Soft Radio-Frequency Identification Sensors: Wireless Long-Range Strain Sensors Using Radio-Frequency Identification
    L. Teng, K. Pan, M.P. Nemitz, R. Song, A.A. Stokes
    Soft Robotics, 6(1), 82-94
  • Integrating Soft Sensor Systems Using Conductive Ink
    L. Teng, K. Jeronimo, T. Wei, M.P. Nemitz, G. Lyu, A.A. Stokes
    Journal of Micromechanics and Microengineering, 28(5), 054001

Soft robot with embedded RFID sensors. Each leg (pneunet) of the fluidically-driven soft robot contains a microfluidic strain sensor: a RFID antenna and a RFID chip. These sensors do not require an external power supply. A change in strain leads to a change in resonance frequency.

Comments are closed.