A benchmarking framework for embodied neuromorphic agents

Nature Machine Intelligence, 2026
Perspective  ·  Published 11 March 2026
Vol. 8, pp. 300–312

A benchmarking framework for embodied neuromorphic agents

Enabling robots to swiftly, robustly and efficiently interact with a dynamic environment remains a key challenge. The robotic community can draw inspiration from the co-adaptation and synergistic interplay between animals' brains and bodies, which underpins embodied intelligence. Soft robots and neuromorphic technology offer a natural solution to such challenges.

Authors

Giulia D'Angelo, Jens E. Pedersen, Taimoor Hassan, Matteo Cianchetti, Josh Bongard, Fumiya Iida, Giacomo Indiveri, Matej Hoffmann, Cecilia Laschi, Chiara De Luca, Chiara Bartolozzi & Elisa Donati

Czech Technical University in Prague (Czech Republic)  ·  Istituto Italiano di Tecnologia, Genoa (Italy)  ·  Technical University of Denmark, DTU Electro (Denmark)  ·  Ghulam Ishaq Khan Institute of Engineering Sciences and Technology (Pakistan)  ·  BioRobotics Institute, Scuola Superiore Sant'Anna, Pisa (Italy)  ·  University of Vermont (USA)  ·  University of Cambridge (UK)  ·  Institute of Neuroinformatics, University of Zurich and ETH Zurich (Switzerland)  ·  National University of Singapore, Advanced Robotics Centre (Singapore)

8 Countries 12 Co-authors

Towards embodied intelligence

Enabling robots to interact swiftly, robustly and efficiently with dynamic environments remains a key challenge. The extraordinary abilities of biological systems lie not only in neural processing, but in the closed-loop interplay between brain, body and environment; what is referred to as embodied intelligence. Soft robots and neuromorphic technology offer a natural solution: by taking biological inspiration beyond the brain, complete sensorimotor loops can be implemented with low-power, material-based and event-driven computation that seamlessly handles the continuous dynamic demands of embodied agents.

"Today's robotic systems can no longer be evaluated solely based on whether they achieve a goal. We also need to consider how quickly they respond, how efficiently they use energy, how well they adapt to changes in the environment, and how effectively they exploit the properties of their own body. We argue that neuromorphic computing and soft robotics together offer a natural solution to these challenges. This is exactly what the proposed benchmarking framework aims to address." — Giulia D'Angelo, CTU FEL Read the CTU FEL press release →

Benchmark metrics

The framework combines traditional robotics metrics (task performance and power consumption) with additional metrics crucial in complex, real-world settings, to quantify the performance of the neuromorphic body–brain dyad. These metrics provide the necessary ingredients to compute figures of merit through multi-objective optimisation, targeting specific application requirements for evaluating the performance of the holistic system:

  • Time performance — execution time (s) and reaction time (s). Time to complete the task and the time required to switch policy in response to environmental changes (reflex <10 ms, reactive <100 ms, adaptive <1 s) for static and dynamic obstacles.
  • Power consumption — total power (W) across sensors, actuation, computation and training
  • System complexity — sensor count and resolution; network parameters including neurons, synapses, and model architecture
  • Task execution — task progression (%), forwards and backwards transfer (%), and agility in performing fast and efficient movements
  • Environmental impact — environmental disruption during operation

Four benchmark task categories of progressively increasing complexity, technology-agnostic and designed to accommodate neuromorphic, event-based, conventional and hybrid systems, challenge the agent's ability to coordinate sensing, computation and actuation across different levels of control:

  • Precision Manoeuvring
  • Adaptive Terrain
  • Parameters Adaptation
  • Grasping and Manipulation
"The goal is to provide the research community with a shared, open, and reproducible foundation for fairly comparing different approaches to robot control under conditions that more closely resemble the real world." — Giulia D'Angelo, CTU FEL Read the CTU FEL press release →

The Active-Braid platform

To enable systematic and reproducible benchmarking, the framework introduces the Active-Braid Platform (AB-Platform) as its dedicated robotic platform. Built on a compliant braided structure that mimics the functionality of muscle layers in muscular hydrostats, it combines modularity and adaptability with precise control over movement and stiffness, making it a natural fit for evaluating the synergies between neuromorphic controllers (brain) and soft body configurations (body). Its peristaltic locomotion, inspired by earthworm movement, supports navigation across challenging terrains as well as grasping and manipulation using the same mechanism.

The platform is open-source, low cost and reproducible: assembly requires only standard tools and widely available materials. All CAD files, Gerber files and the complete list of components are available at github.com/ActiveBraid/ActiveBraidCrawler, providing everything required for replication and benchmarking.

"What is valuable about this work is that it shows the need to evaluate robotic systems as a whole. Intelligence does not arise only in the control architecture, but in the constant interplay between computation, body, and environment — an area we have long been pursuing at FEL." — Matěj Hoffmann, CTU FEL Read the CTU FEL press release →

Press coverage
CTU FEL Press Release (English)

Cite: D'Angelo, G., Pedersen, J.E., Hassan, T., et al. A benchmarking framework for embodied neuromorphic agents. Nat Mach Intell 8, 300–312 (2026). https://doi.org/10.1038/s42256-026-01197-w