PhysicalAI-AV dataset: License clarification for academic use: shared HPC storage and publishing results

Hi, and thank you for releasing the PhysicalAI-Autonomous-Vehicles dataset and the Alpamayo ecosystem.

I am an academic researcher at Linköping University, Sweden, planning to evaluate reasoning VLA models, including Alpamayo, on this dataset. While setting this up, a few questions about the NVIDIA Autonomous Vehicle Dataset License came up that are likely relevant to many academic users. I could not find answers on the dataset card or in the GitHub issues.

1. Shared copy on academic HPC systems
We asked our national HPC centre (NAISS, Sweden) to host a shared copy for its users. They declined, because under Section 2 (Authorized Users / Affiliates) and Section 4.6 (no hosting or making the Dataset available to others), users from other institutions would not qualify. As a result, every group has to download and store its own copy of the 133 TB dataset.

  • Would NVIDIA allow an academic HPC centre to host one copy, accessible only to users who have each accepted the license on Hugging Face themselves?
  • If not, can individual licensees store the dataset in access-restricted project storage on such a third-party academic system? Section 2 refers to access “from your secure network”.

2. Publishing evaluation results
Section 3 lists “results of benchmarking … or performance data relating to the Dataset” as Confidential Information. Can academic papers report evaluation results on the dataset (e.g. on the clips listed in alpamayo-recipes) and show a small number of example frames as figures? NVIDIA’s public challenges and existing publications suggest this is intended to be allowed, but an explicit confirmation would help.

3. Scope of Purpose
Section 1 limits use to development “using NVIDIA technology”. Is evaluating non-NVIDIA models, for example as baselines compared against Alpamayo, within this scope?

A short clarification here, or an FAQ for academic users on the dataset card, would help researchers and HPC centres a lot. I have also opened a documentation request on GitHub: [GitHub issue link]

Thanks!

2 Likes