Open Navigation's robotics benchmark tests full autonomy on real workloads
Run your own robotics compute test with 9 sensor streams and VLM workload.
Open Navigation's latest release, the 'Nav2 Robotics Workload Benchmark,' is a significant departure from synthetic microbenchmarks. It tests compute platforms by running a complete autonomous stack performing forklift pick-and-place operations inside a simulated 180,000 sqft warehouse. The benchmark ingests 9 sensor streams, including 3D LIDAR and depth cameras, while simultaneously running a heavy vision-language model (VLM) workload—mimicking real-world robotics demands. Fully Dockerized and open-source, it allows teams to easily run it on their own hardware and compare results.
The initial benchmark results compare AMD Strix Halo, NVIDIA Jetson Thor, and Orin AGX, with full findings published in a technical report. The release also sparked discussion about AMD's new Kria AI platform as a potential lower-cost alternative to costly Jetson modules. Open Navigation invites community contributions to grow the benchmark over time, making it a living resource for evaluating robot computers. For any team designing autonomous robots, this tool provides a far more accurate picture of real-world performance than isolated tests ever could.
- Benchmark runs a full autonomy stack for forklift pick-and-place in a 180,000 sqft warehouse environment.
- Processes 9 sensor streams (3D LIDAR, depth cameras) with a heavy VLM workload for realistic compute testing.
- Fully open-source, Dockerized, and tested on AMD Strix Halo, NVIDIA Jetson Thor, and Orin AGX.
Why It Matters
Enables robotics teams to benchmark hardware with real workloads, leading to more informed compute platform decisions.