Year: 2026
Type: editorial
Abstract: framework and highlight the value of uncertainty modelling for robotic perception in adverse visual conditions.This article provides a model compression for practical deployment. The second contribution surveys the development of model compression techniques and their role in deploying large and domain-specific models under practical hardware constraints. Large models require substantial training e[ort, memory, and computational capacity, making e[icient inference di[icult on embedded and edge platforms. The review examines quantisation, pruning, low-rank decomposition, and knowledge distillation, outlining their principles, recent progress, and deployment implications. By organising past and current approaches while identifying future research directions, the survey provides a technical reference for reducing model size and computational demand without abandoning application-level performance.This article is about energy-e[icient embodied spiking neural networks. The third contribution presents SNN4Agents, a framework for designing energy-e[icient embodied spiking neural networks for autonomous agents. Spiking neural networks can exploit sparse event-based computation, but systematic optimisation procedures for deployment remain underdeveloped. SNN4Agents combines weight quantisation, timestep reduction, and attention-window reduction to improve energy e[iciency, memory use, and latency while maintaining accuracy. In eventbased car recognition on the NCARS dataset, the framework achieves 84.12% accuracy together with 68.75% memory savings, a 3.58× speed-up, and a 4.03× improvement in energy e[iciency relative to the referenced state-of-the-art method. These results illustrate the potential of coordinated optimisation for practical neuromorphic autonomy.This article is about mapless navigation with self-supervised cognitive maps. The fourth contribution addresses mapless navigation in unknown environments using a shallow architecture trained through a local learning rule. Deep reinforcement learning methods can achieve strong navigation performance, but multi-layer networks, complex reward objectives, and backpropagation increase memory and computational demands. The proposed method combines novelty-based exploration with random walks, reducing dependence on manually defined rewards and supporting greater agent autonomy. It reaches performance comparable to a deep Q-network in goal-reaching accuracy and path length while avoiding error backpropagation. Its shallow structure and local learning mechanism make it suitable for edge neuromorphic processors and resource-constrained robots.Together, the contributions show that embodied neuromorphic AI is not defined by a single algorithm or device. Progress depends on the joint design of sensing, representation, learning, model optimisation, hardware, and control. Robust perception under uncertainty, compressed models, e[icient spiking networks, and local self-supervised learning are complementary steps toward autonomous agents that can operate for longer periods, react in real time, and adapt with limited resources. Future work should strengthen integration across sensors, processors, and actuators; establish clearer programming and benchmarking frameworks; and evaluate complete systems in realistic, changing environments. Such coordinated e[orts can help neuromorphic robotics move from specialised demonstrations toward dependable agents that interact naturally and safely with society.
Source: Frontiers in Robotics and AI
Authors: Jorge Dias, Fakhreddine Zayer, Mario Molinara
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,353
OpenAlex ID: W7203771855
Cites:
Cited by:
Related to: