Neuromorphic force-control in an industrial task: validating energy and latency benefits

TL;DR

This work demonstrates energy- and latency-efficient neuromorphic control for a real-world industrial peg-in-hole task. A spiking neural network trained with reinforcement learning in simulation is ported to the Loihi neuromorphic chip and deployed on a KUKA arm, with sim2real techniques bridging the gap to reality. Results show competitive inference latency and an order-of-magnitude energy savings compared with conventional edge hardware, achieving 100% insertion success on real hardware after domain randomization and system identification. The study provides a concrete proof-of-concept for neuromorphic controllers in robotics and outlines practical considerations for integrating spike-based computation with real actuators.

Abstract

As robots become smarter and more ubiquitous, optimizing the power consumption of intelligent compute becomes imperative towards ensuring the sustainability of technological advancements. Neuromorphic computing hardware makes use of biologically inspired neural architectures to achieve energy and latency improvements compared to conventional von Neumann computing architecture. Applying these benefits to robots has been demonstrated in several works in the field of neurorobotics, typically on relatively simple control tasks. Here, we introduce an example of neuromorphic computing applied to the real-world industrial task of object insertion. We trained a spiking neural network (SNN) to perform force-torque feedback control using a reinforcement learning approach in simulation. We then ported the SNN to the Intel neuromorphic research chip Loihi interfaced with a KUKA robotic arm. At inference time we show latency competitive with current CPU/GPU architectures, and one order of magnitude less energy usage in comparison to state-of-the-art low-energy edge-hardware. We offer this example as a proof of concept implementation of a neuromoprhic controller in real-world robotic setting, highlighting the benefits of neuromorphic hardware for the development of intelligent controllers for robots.

Figures (6)

• Figure 1: Learning and inference approach. HIL refers to 'Hardware-In-Loop". An in-depth look at the third step can be found in Figure \ref{['fig:FlowChart']}.
• Figure 2: Neurorobotics simulation setup.
• Figure 3: Demonstrator setup with a KUKA IIWA 7 R800 robot and attached cylindrical peg, black target box with hole, and Kapoho Bay containing two Loihi chips (on the tabletop).
• Figure 4: An overview of the communication mechanism between the high-level neuromorphic controller running on Loihi and the low level controller operating on the FRI.
• Figure 5: Learning curves showing the results of training with 20 random seeds, the mean performance and the standard deviation.