Exploring Neuromorphic Concepts for Advancing Robotic Perception: Towards Embodied Neuromorphic AI

Talk Abstract

In recent years, there has been a growing interest in developing robotic systems with enhanced perceptual capabilities, enabling them to interact intelligently with their environments. Traditional approaches to robotic perception often rely on algorithms that process sensory information in a step-by-step manner, which can be computationally expensive and lack the flexibility exhibited by biological systems. To address these limitations, the emerging field of neuromorphic engineering offers a promising alternative by drawing inspiration from the structure and function of the human brain. This abstract presents an investigation into the utilization of neuromorphic concepts for robotic perception, aiming to bridge the gap between artificial intelligence and embodied cognition. By leveraging the inherent parallelism and efficiency of neuromorphic hardware, we aim to develop robotic systems that can perceive and understand their surroundings in a manner closer to biological organisms. Our research focuses on the integration of neuromorphic hardware with state-of-the-art machine learning algorithms, allowing robots to process sensory data in real-time and extract meaningful information with minimal computational resources. We propose a hybrid approach that combines the strengths of traditional computer vision techniques with neuromorphic architectures, enabling robots to adapt and learn from their experiences while exhibiting robustness and energy efficiency. The envisioned applications of our work span a wide range of domains, including autonomous navigation, object recognition, scene understanding, and human-robot interaction. By utilizing neuromorphic principles, we aim to empower robots with the ability to perceive and interpret their environment more efficiently, leading to improved decision-making, adaptability, and overall performance. In this talk, we will present our ongoing research efforts in developing neuromorphic-inspired algorithms for robotic perception. We will discuss the design considerations, experimental methodologies, and results obtained thus far, highlighting the advantages and challenges associated with adopting a neuromorphic approach. Furthermore, we will explore the potential impact of our findings on the development of next-generation robotic systems and their role in shaping the future of embodied neuromorphic AI. Keywords: neuromorphic engineering, robotic perception, embodied cognition, machine learning, computer vision, parallel computing, adaptive systems, intelligent robotics, neuromorphic hardware, artificial intelligence.

Dr. Fakhreddine Zayer

Vote of Thanks

Solving Robotic Manipulation with Sparse Reward Reinforcement Learning and Hyperdimensional computing

Talk Abstract

Brief Description on the talk: Learning with sparse rewards is usually inefficient in Reinforcement Learning (RL). Hindsight Experience Replay (HER) has been shown an effective solution to handle the low sample efficiency that results from sparse rewards by goal relabeling. However, the HER still has an implicit virtual-positive sparse reward problem caused by invariant achieved goals, especially for robot manipulation tasks. To solve this problem, novel model-free continual RL algorithms, such as Relay-HER (RHER) have been proposed. The latter is based on the decomposition and rearrangement of the original long-horizon task into new sub-tasks with incremental complexity. Subsequently, a multi-task networks are designed to learn the sub tasks in ascending order of complexity. To solve the virtual-positive sparse reward problem, this talk will target Random-Mixed Exploration Strategies, in which the achieved goals of the sub-task with higher complexity are quickly changed under the guidance of the one with lower complexity.

Dr. Benjamin Tee

Speaker biography

Dr. Benjamin C.K. Tee is appointed President’s Assistant Professor in Materials Science and Engineering Department at the National University of Singapore (NUS). He obtained his PhD (EE) at Stanford University, and was a Singapore-Stanford Biodesign Global Innovation Postdoctoral Fellow in 2014. He has developed and patented several award-winning electronic skin sensor technologies. He is an MIT TR35 Innovator (Global) in 2015 and Singapore National Research Foundation (NRF) Fellow. His research group aims to develop technologies at the intersection of materials science, mechanics, electronics and biology, with a focus on sensitive electronic skins that has tremendous potential to advance global healthcare technologies in an increasingly Artificial Intelligence (AI) and Robotics driven era.

Vote of Thanks

Coffee break

Vote of Thanks

Enabling Neuromorphic Perception and Control with Virtual Neurorobotics

Talk Abstract

Neuromorphic computing bears huge potential for many applications in robotics, with some of the main advantages being low-latency processing, support for online learning, and extremely low energy consumption. As more and more neuromorphic computing platforms are becoming available, the main challenge today lies in interfacing them with robotic systems and, most importantly, the development of new algorithms. This opens up a vast design space that is impossible to explore with traditional experiments on physical robots. In this talk, we highlight how virtual experiments performed in simulation enable the efficient design of neuromorphic perception and control algorithms for robotics. We provide an overview of the Neurorobotics Platform (NRP), our tool set for the modeling and simulation of brain-derived systems and neuromorphic algorithms. It is highly scalable and can interface with state-of-the-art neuromorphic processors. We further demonstrate how the NRP can be used to implement models and algorithms from computational neuroscience in robotics and present results from our research on neuromorphic computing in reinforcement learning and automotive sensor data processing.

Eng. Florian Walter

Speaker biography

Florian Walter received his Master’s degree in informatics with high distinction from Technische Universität München. During his Master studies, he completed an internship in the automotive industry and was a visiting student researcher in the Artificial Intelligence Laboratory at Stanford University where he worked in the field of online trajectory generation for robotics. In August 2014, Florian joined the HBP SP10 Neurorobotics research group as research assistant where he is now working in the field of neurobiological learning methods for robotics.

Vote of Thanks

Intelligent Memory for Efficient Hardware Accelerator

Talk Abstract

Memory architecture and design have been critical for digital systems to achieve ample storage, low latency, fast access time, and energy efficiency, especially for battery-operated devices. The increase of data generated by many devices such as mobile, sensors, communications, and security not only increased the requirements on memory capacity but also increased the challenges on memory access and energy. The memory interface has limited throughput and high latency, which has not been scaling at the same rate as data size or processing speed; this limits the performance of accessing the data, which refer to as the memory wall. In addition to the negative impact on latency and performance, large data movement results in high energy consumption.  Research has been focusing on elevating the memory wall issue by engineering more memory hierarchy and increasing local on-chip memory. This has partially reduced the timing issue but did not address the high leakage and active energy consumption. It is estimated that more than 60% of energy spent on most computing platforms is spent on data movements and memory access. The new era of big data and artificial intelligence-based applications has increased the urgency to solve memory capacity, data movement energy, and memory wall issues. Some solutions have brought processing into centralized cloud computing, with high performance and large memory hardware capacity available. However, this brought a new challenge to communications, privacy, security, and latency, especially for real-time applications. The goal of this lecture is to highlight the after mentioned challenges and to present a new paradigm of computing beyond von Neuman's architecture to enable processing as close to the data source as possible. This includes in-memory computing, near memory computing architecture. Both existing and emerging memory technologies will be explored. Since the new computing paradigm is more data-centric than traditional processing-centric, the traditional single architecture for all applications is not feasible, but rather a domain-specific architecture and hardware solutions need to be adopted.  Popular high computing functions such as Query, MAC, hamming distance, and image compression will be presented as an example of in-memory hardware accelerators.

Baker Mohammad

Vote of Thanks

Brain-Inspired Hyperdimensional Computing for Robotic Manipulation

Talk Abstract

Hyperdimensional computing (HDC) architectures in machine learning have gained unprecedented popularity in recent years due to their simplicity, robustness, and computational efficiency. Moreover, when compared with the deep neural networks, that are both computationally expensive and require exhaustive training rounds, HDC emerges as an excellent alternative that offers a highly competitive performance with extremely lower computational complexity as compared to the neural networks. Due to the optimal trade-off between computational efficiency and recognition performance, HDC showcases a strong potential to be deployed on a resource-constrained devices that are used in wide variety of applications related to robot sensing, perception, and manipulation. In this talk, we will discuss some of the advances in brain inspired HDC architectures to solve robotic manipulation tasks. More specifically, we will discuss some frameworks that improve the performance of soft-actor critic (SAC)-based reinforcement learning algorithms for grasping moving objects using HDC architectures, that are trained using trajectory-based contrastive unsupervised learning. Through simulation and real-world experiments, we will see how HDC models and trajectory-based learning can overcome the inherent limitations of sample inefficiency for the SAC-based RL agents towards grasping moving objects.

Dr. Taimur Hassan

Vote of Thanks

Coffee break

Vote of Thanks