Event-Driven Sensing for a Humanoid Robot

Event: Event-Driven Sensing Workshop 2016 · Duration: 31 min · ▶ Watch on YouTube

Abstract

This presentation delves into the development and application of event-driven sensing technologies for humanoid robots, specifically focusing on the iCub platform. The speaker emphasizes the importance of autonomy, low power consumption, and real-time decision-making, drawing parallels with the efficiency of biological systems. Novel event-based approaches are showcased for both tactile sensing, utilizing POSFETs, and vision, employing ATIS sensors. The talk demonstrates how these event-driven methods can be leveraged for tasks such as optical flow estimation, corner detection, independent motion detection, and vergence control, offering significant advantages over traditional frame-based techniques in terms of data efficiency, latency, and robustness to environmental changes.

Speakers

  • Chiara Bartolozzi — Istituto Italiano di Tecnologia (IIT)

Talks (1)

  • 00:00:00 — Chiara Bartolozzi: Event-Driven Sensing for a Humanoid Robot
    • This talk explores the integration of event-driven vision and touch sensors with the iCub humanoid robot, highlighting the benefits of this approach for achieving greater autonomy, power efficiency, and real-time performance in robotics, drawing inspiration from biological systems.

Key Takeaways

  • Event-driven sensing is crucial for achieving high autonomy and power efficiency in humanoid robots by processing information only when changes occur, mimicking biological systems.
  • Integrating event-driven touch sensors (like POSFETs) on robots like iCub can significantly reduce data bandwidth and computation due to the sparse and localized nature of tactile signals.
  • Novel algorithms are required to adapt traditional computer vision tasks (e.g., optical flow, corner detection) to continuous event streams, exploiting temporal information for real-time, low-latency processing.
  • Event-based vision systems demonstrate superior performance in challenging conditions, such as varying illumination and tracking fast-moving targets, without the need for normalization, unlike frame-based cameras.
  • The combination of event-driven vision and touch, along with neuromorphic processing, offers a powerful platform for developing biologically inspired, multi-sensory robot perception and control.

Methods / Models / Datasets Mentioned

  • DVS
  • POSFET
  • SpiNNaker
  • TrueNorth
  • AER (Address-Event Representation)
  • Harris Corner Detection
  • Phase-Shift Model
  • Gabor filters
  • Particle Filter
  • Hough Transform

Topics

Event-driven sensing · Humanoid robots · iCub · Neuromorphic computing · Tactile sensing · Event-driven vision · Optical flow · Corner detection · Vergence control · Low-power robotics · Biological inspiration

Notes

Open for commentary — connections to other work, critiques, follow-up reading.