25 November 2020 in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
Neuromorphic engineering promises the deployment of low latency, adaptive and low power systems that can lead to the design of truly autonomous artificial agents. However, many building blocks for developing a fully neuromorphic artificial agent are still missing. While neuromorphic sensing, perception, and decision-making building blocks are quite mature, the ones for motor control and actuation are lagging behind. In this paper we present a closed-loop motor controller implemented on a mixed-signal analog/digital neuromorphic processor which emulates a spiking neural network that continuously calculates an error signal from the desired target and the feedback signals. The system uses population coding and recurrent Winner-Take-All networks to encode the signals robustly. Recurrent connections within each population are used to speed up the convergence, decrease the effect of mismatch and improve selectivity. The error signal computed in this way is then fed into three additional populations of spiking neurons which produce the proportional, integral and derivative terms of classical controllers exploiting the temporal dynamics of the network synapses and neurons. To validate this approach we interfaced this neuromorphic motor controller with an iCub robot simulator. We tested our spiking controller in a single joint control task for the robot head yaw. We demonstrate the correct performance of the spiking controller in a step response experiment and apply it to a target pursuit task.
2 November 2022 in Research Square
Current low latency neuromorphic processing systems, and future ones based on ultra-low power mixed-signal circuits in advanced technology nodes and memristive nano-scale devices, hold great potential for developing autonomous artificial agents. However, the variable nature and low precision of the underlying hardware substrate pose severe challenges for robust and reliable performance. To address these challenges, we adopt hardware-friendly processing strategies based on brain-inspired computational primitives, such as triplet spike-timing dependent plasticity, basal ganglia inspired disinhibition, and cooperative-competitive networks. We demonstrate this approach by presenting an example of robust real-time motor control using a hardware spiking neural network implemented on a mixed-signal neuromorphic processor, trained to learn the inverse kinematics of a 2-joint robotic arm. The final system is able to perform low-latency control robustly and reliably using noisy silicon neurons. The spiking neural network, trained to control two joints of the iCub robot arm simulator, performs a continuous target reaching task with 97.93% accuracy, 33.96 ms network latency, 102.1 ms system latency and with an estimated power consumption of 26.92 µW. This work provides insights into how specific computational primitives used by real neural systems can be applied to neuromorphic computing for solving real-world engineering tasks. It represents a milestone in the design of end-to-end spiking robotic control systems, relying on event-driven sensory encoding, neuromorphic processing, and spiking motor control.
2 September 2025 in IEEE Robotics and Automation Letters
Inverse dynamics maps the joint space of a robotic system to the torque space, but traditional methods for solving it rely on direct physical modeling of robots, which is difficult or even impossible due to high nonlinearity and external disturbance. Data-driven methods address these issues by learning the complex and unknown inverse dynamics. One of the key features of inverse dynamics is temporal dependency because robot movements are continuous. While recent Artificial Neural Network (ANN) methods leverage this to improve prediction accuracy, their high computational cost limits use in edge computing scenarios. Neuromorphic computing based on Spiking Neural Network (SNN) features extremely low power consumption. However, existing neuromorphic approaches remain limited to 2 degrees-of-freedom (DoFs) due to insufficient temporal modeling and low scalability. This paper presents the first neuromorphic approach capable of accurately learning the inverse dynamics of up to 7 DoF robotic manipulators. The method utilizes a liquid state machine (LSM) with temporally structured inputs and a novel decoding strategy to better model the temporal dependency in robotic motion. To account for the distinct characteristics of each joint while enhancing the model's scalability, we adopt a multi-LSM structure, enabling individualized modeling for each joint. Our neuromorphic method outperforms state-of-the-art approaches by at least 14.6% while using only 40% of their power consumption. It achieves the highest prediction accuracy across four different robotic platforms and is further validated on a UR5 robotic manipulator in a trajectory tracking task.