Session
Advanced Technologies Research & Academia 1
Location
Salt Palace Convention Center, Salt Lake City, UT
Abstract
We present an end-to-end pipeline for deploying neural networks on low SWaP neuromorphic hardware to control a CubeSat like robot. As a test case, we use a reinforcement learning (RL)-trained Artificial Neural Networks (ANNs) on neuromorphic hardware by converting them into spiking Sigma-Delta Neural Networks (SDNNs) for controlling the NASA Astrobee free-flying robot, similar to a previously space-validated controller demonstrated on robotic hardware. We demonstrate that an ANN trained entirely in simulation can be transformed into an SDNN compatible with Intel’s Loihi 2 neuromorphic architecture, enabling low-latency and energy-efficient inference. The SDNN is deployed on Loihi 2, then evaluated in NVIDIA’s Omniverse Isaac Lab simulation environment for closed-loop control of Astrobee’s motion. We compare execution performance between GPU and Loihi 2. The results highlight the feasibility of neuromorphic platforms for robotic control and establish a pathway toward energy-efficient, real-time neuromorphic computation for future space and terrestrial robotics applications.
Power constraints are a critical consideration for many robotic applications, particularly in space and mobile environments. While data-driven learning on GPUs has pushed substantial progress in robotics, the associated energy demands can hinder their deployment in power-sensitive applications. This paper explores a path toward lower-power robotic control by leveraging neuromorphic hardware. Results demonstrate an SDNN running on Loihi 2 for controlling the Astrobee is 20x more energy efficient with 2x throughput compared to running on a GPU with only a small cost to accuracy.
Beyond small free-flying platforms such as Astrobee, many real-world missions require robust, low-power control systems capable of sustained operation under strict energy budgets. Space exploration is particularly constrained with respect to onboard computing resources due to environmental challenges, including radiation exposure, as well as size, weight, power, and cost (SWaP-C) limitations. Consequently, current radiation-hardened processors, while reliable for long-duration missions, provide limited computational performance compared to modern terrestrial hardware.
Several past space robotic platforms encountered limitations due to actuator degradation, restricted onboard computation, and power exhaustion, conditions under which neuromorphic RL-based control could provide tangible benefits. For example, NASA’s Kepler mission experienced reaction wheel degradation that reduced pointing accuracy and ultimately ended its primary mission. More adaptive torque-management policies learned through reinforcement learning and executed at low power on neuromorphic hardware could have mitigated reaction-wheel loading and prolonged operational life. Continuous, low-power neuromorphic inference may also enable more resilient attitude and propulsion control during sensor anomalies without exceeding spacecraft power constraints.
These considerations motivate the development of control pipelines that integrate data-driven learning with low-power neuromorphic execution. By demonstrating an ANN-to-SDNN conversion pipeline for RL-based robotic control and validating it in a high-fidelity simulation environment, this work supports the development of autonomous systems capable of long-term, energy-efficient operation, addressing critical resource constraints with neuromorphic platforms such as Loihi 2.
Document Type
Event
Autonomous Reinforcement Learning Astrobee Control With Low SWaP Neuromorphic Hardware
Salt Palace Convention Center, Salt Lake City, UT
We present an end-to-end pipeline for deploying neural networks on low SWaP neuromorphic hardware to control a CubeSat like robot. As a test case, we use a reinforcement learning (RL)-trained Artificial Neural Networks (ANNs) on neuromorphic hardware by converting them into spiking Sigma-Delta Neural Networks (SDNNs) for controlling the NASA Astrobee free-flying robot, similar to a previously space-validated controller demonstrated on robotic hardware. We demonstrate that an ANN trained entirely in simulation can be transformed into an SDNN compatible with Intel’s Loihi 2 neuromorphic architecture, enabling low-latency and energy-efficient inference. The SDNN is deployed on Loihi 2, then evaluated in NVIDIA’s Omniverse Isaac Lab simulation environment for closed-loop control of Astrobee’s motion. We compare execution performance between GPU and Loihi 2. The results highlight the feasibility of neuromorphic platforms for robotic control and establish a pathway toward energy-efficient, real-time neuromorphic computation for future space and terrestrial robotics applications.
Power constraints are a critical consideration for many robotic applications, particularly in space and mobile environments. While data-driven learning on GPUs has pushed substantial progress in robotics, the associated energy demands can hinder their deployment in power-sensitive applications. This paper explores a path toward lower-power robotic control by leveraging neuromorphic hardware. Results demonstrate an SDNN running on Loihi 2 for controlling the Astrobee is 20x more energy efficient with 2x throughput compared to running on a GPU with only a small cost to accuracy.
Beyond small free-flying platforms such as Astrobee, many real-world missions require robust, low-power control systems capable of sustained operation under strict energy budgets. Space exploration is particularly constrained with respect to onboard computing resources due to environmental challenges, including radiation exposure, as well as size, weight, power, and cost (SWaP-C) limitations. Consequently, current radiation-hardened processors, while reliable for long-duration missions, provide limited computational performance compared to modern terrestrial hardware.
Several past space robotic platforms encountered limitations due to actuator degradation, restricted onboard computation, and power exhaustion, conditions under which neuromorphic RL-based control could provide tangible benefits. For example, NASA’s Kepler mission experienced reaction wheel degradation that reduced pointing accuracy and ultimately ended its primary mission. More adaptive torque-management policies learned through reinforcement learning and executed at low power on neuromorphic hardware could have mitigated reaction-wheel loading and prolonged operational life. Continuous, low-power neuromorphic inference may also enable more resilient attitude and propulsion control during sensor anomalies without exceeding spacecraft power constraints.
These considerations motivate the development of control pipelines that integrate data-driven learning with low-power neuromorphic execution. By demonstrating an ANN-to-SDNN conversion pipeline for RL-based robotic control and validating it in a high-fidelity simulation environment, this work supports the development of autonomous systems capable of long-term, energy-efficient operation, addressing critical resource constraints with neuromorphic platforms such as Loihi 2.
