Closed-Loop Neuromorphic Benchmarks.

Area of Science:

• Neuroscience
• Robotics
• Computer Engineering

Background:

• Evaluating neuromorphic hardware, especially in closed-loop systems where output influences future input, presents significant challenges.
• Closed-loop applications are a primary use case for neuromorphic hardware, necessitating effective evaluation methods.

Purpose of the Study:

• To develop a flexible methodology for creating closed-loop benchmarks for neuromorphic hardware.
• To demonstrate the methodology's utility in assessing motor control performance under complex conditions.

Main Methods:

• A hybrid approach combining real physical embodiment with "minimal" simulation to create closed-loop benchmarks.
• Development of novel benchmarks focusing on motor control for systems with unknown external forces and up to 15 interacting joints.

Main Results:

• The proposed methodology facilitates robust real-world performance evaluation while retaining simulation's practical advantages.
• An error-driven learning rule demonstrated consistent improvement in motor control performance across diverse closed-loop simulations.

Conclusions:

• The presented methodology offers a flexible and effective way to benchmark neuromorphic hardware in closed-loop scenarios.
• This approach enables researchers to identify specific task domains where particular hardware demonstrates superior performance, particularly in motor control applications.