Open NavigationRobotics Workload Benchmark

Contact

Talk to us, get your platform included, or add your own results

Four different reasons you might be on this page. Pick the one that is yours.

Who runs this

This benchmark is built and maintained by Open Navigation LLC, the maintainers of Nav2. That is the open-source autonomy framework the workload itself is built on, and the one a large share of production mobile robots run today.

It is an independently managed, vendor-agnostic benchmark. No hardware vendor controls the workload definition, the metrics, the analysis, or what gets published. The reason we can say that and expect to be believed is not the claim itself. The pipeline, raw logs, and the full analysis results are all in the open, so any metric here can be recomputed or contested by anyone who disagrees - all coming from a trusted source.

The work is made possible with support from benchmarked vendors. Compute OEMs pay for engineering support to maintain the benchmark, but does not buy a result and results unflattering to any platform are published alongside the rest.

opennav.org GitHub repository

Commercial support

Open Navigation provides commercial engineering support for teams building on ROS 2 and Nav2. If you are choosing hardware using this benchmark, we can also help you validate the choice and other architectural, hardware, ROS, Nav2, and autonomy needs you may have.

info@opennav.org Open Navigation services

Test your platform yourself

The pipeline is public and designed to be extended and tested yourself for your own platform(s) and software. If you run it on hardware that is not here and want to publish the results, lets talk.

Full instructions for adding a platform

The checklist
  1. Implement detect_platform() detection, aHARDWARE_PROFILES entry, and a GpuMetrics subclass inopennav_benchmark_pipeline/scripts/hardware_platforms.py so your GPU, NPU, clock, and power counters are read natively.
  2. Add docker/ai_workload/<platform>/Dockerfile andprofile.sh, servingggml-org/gemma-4-31B-it-GGUF:Q4_K_M behind a llama.cpp server, with whatever GPU runtime and device passthrough flags your platform needs.
  3. Measure sensor-driver load on your hardware with real lidar and depth cameras, and put the coefficients in your HARDWARE_PROFILES entry.
  4. Run run_benchmark.sh for 900 seconds in each power mode you want represented, with the simulation on a separate machine over wired Ethernet.
  5. Open a pull request with the fullopennav_benchmark_logs/<category>/<platform>/ tree.
What we need to publish a result
  • A complete system_metrics.json covering the whole run.
  • The full ros/ log directory. Mission counts, control-loop misses, planner cycle times, and VLM outcomes are all parsed from it.
  • The configured TDP, plus firmware and BIOS versions.
  • The simulation machine's specification, so we know the sensor load was actually delivered.
  • Any vendor relationship, disclosed. A result from a vendor is welcome; an undisclosed one is not.

Submit platform results