Why use Loihi 2 instead of a GPU for robot control?

Research on sensor fusion with Loihi 2 found it over 100 times more energy efficient than a CPU and nearly 30 times more efficient than a GPU for the workloads tested, which matters most for power-constrained platforms like drones and free-flying robots.

June 22, 2026

Neuromorphic Robotics with Loihi 2 and ROS2: The Research So Far

Q: Has Loihi 2 been integrated with ROS2?

Yes, in research settings. Published work includes an Astrobee free-flying robot reinforcement-learning policy converted to a spiking neural network for Loihi 2, hexapod locomotion control synchronized with ROS, and a multi-robot TurtleBot3 capture-the-flag demo using ROS 2 Humble with a centralized policy running on a Loihi Kapoho Bay chip.

Q: Can I deploy to Loihi 2 without Intel's archived Lava SDK?

Yes. NeuroCUDA validates a Loihi 2 simulator target against Intel's published neuron equations to zero spike deviations across 256,000 comparisons, and exposes it as a deploy target without requiring Lava, which Intel has since archived.

Loihi 2 has been connected to ROS and ROS2 in at least four published research efforts. None of them are a package you can install today.

"Neuromorphic robotics" - running spiking neural networks on neuromorphic chips like Intel's Loihi 2 to control real robots - has moved from pure simulation to a string of real hardware demonstrations over the past few years. The pattern across nearly all of them: one robot, one paper, no reusable package.

What's been published

Project	Robot / platform	Approach
Astrobee RL-to-SNN pipeline	Astrobee free-flying robot (NRL + Intel Labs)	RL policy trained in NVIDIA Isaac Lab, converted to SNN for Loihi 2
Astrocyte-modulated CPG	Hexapod robot	Loihi Kapoho Bay synchronized with ROS at 100 Hz for locomotion
S2Act	Multiple TurtleBot3 Burger robots	Centralized policy on Loihi (Kapoho Bay), ROS 2 Humble + motion capture
On-chip multi-component control	Vision-based robot control (research platform)	Spiking state machine orchestrating multiple networks entirely on-chip

The Astrobee work is notable for its pipeline structure: train a reinforcement-learning policy as a normal ANN (where training is easy and sim-to-real transfer is well understood), then convert it to a spiking network only at the deployment stage - rather than training a spiking network from scratch. That conversion-not-training approach is the same philosophy behind ANN-to-SNN compilers generally, including NeuroCUDA.

Why energy efficiency keeps coming up

A Drexel University-led sensor fusion study running on Loihi 2 measured over 100x energy efficiency versus a CPU and nearly 30x versus a GPU on the workloads tested. That gap is the entire argument for neuromorphic robotics: free-flying robots, drones, and other power-constrained platforms can't carry a GPU's power budget, but they can carry Loihi 2's milliwatt-range draw - if the software pipeline to get a model onto the chip exists.

The common thread: ROS2 as the integration layer, not the SNN runtime

In every project above, ROS2 (or ROS1) isn't running the spiking network itself - it's the layer that gets sensor data to the neuromorphic chip and motor commands back out. The hexapod project had to explicitly handle the timing mismatch: Loihi runs asynchronously, ROS runs on a fixed-rate loop, so the ROS node had to block and wait for the chip rather than the reverse. That kind of synchronization code gets rewritten per project today because there's no shared package handling it.

What's still missing for production use

No installable package. Every project above is research code tied to one robot and one paper, not something you pip install or colcon build against your own robot.
Lava's archival. Intel's lava-nc repositories, the primary software path to Loihi 2, are now archived with no announced successor SDK - a real gap for anyone trying to start fresh, covered in our piece on Lava's archival.
No standard ROS2 message set for spike data. Each project invents its own representation for spikes and neuromorphic sensor data rather than sharing one.

NeuroCUDA's ROS2 package addresses the second and third points directly: it doesn't depend on Lava, instead validating a Loihi 2 simulator target against Intel's published neuron equations to zero spike deviations across 256,000 comparisons, and it defines typed messages - SnnDetection, SnnSpikeEvent, SnnStatus - meant to be reused across models rather than invented per project. See the architecture page for the full message definitions, and the deployment walkthrough for the launch-file workflow.

Where this fits

For the broader survey of ROS2-and-SNN projects beyond just Loihi 2, see ROS2 and Spiking Neural Networks: What Exists, What Doesn't. New to ROS2 itself? Start with What Is ROS2?

Frequently asked questions

Has Loihi 2 been integrated with ROS2? Yes, in research: the Astrobee pipeline, hexapod locomotion, and a TurtleBot3 multi-robot demo on ROS 2 Humble.

Why use Loihi 2 instead of a GPU? Measured energy efficiency over 100x better than a CPU and nearly 30x better than a GPU in published sensor-fusion benchmarks.

Can I target Loihi 2 without Lava? Yes - NeuroCUDA's Loihi 2 simulator target is validated independently of Lava, which Intel has archived.