Several policies download assets from Hugging Face Hub (for example, SmolVLA, Pi0, and other Hub-backed models).
If HF_TOKEN is not set, the backend logs a warning and Hub access is
unauthenticated.
Set HF_TOKEN for any workflow that depends on Hugging Face-hosted assets.
Without a token, model downloads may fail (for example, due to anonymous rate
limits or access restrictions on gated/private repositories).
For model download in Physical AI Studio, use a token with read-only Hub access:
- Required:
Readpermission for model repositories. - Not required:
Writeor admin permissions. - If using gated/private models: the token must belong to a Hugging Face account that has been granted access to those repositories.
If you use a fine-grained token, grant read access to the specific model repos you plan to train from.
- Sign in to huggingface.co.
- Open Settings -> Access Tokens.
- Create a new token.
- Set permissions to read-only model access (see required permissions above).
- Copy the token value.
Set HF_TOKEN in the environment used by the backend.
Add the token to application/backend/.env:
HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxThen start the backend as usual:
cd application/backend
./run.shAdd the token to application/docker/.env:
HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxThen run Docker Compose as usual:
cd application/docker
docker compose up- Start a training job for a Hub-backed policy (for example, SmolVLA or Pi0).
- Confirm there is no warning about missing
HF_TOKEN.
- Never commit real tokens to source control.
- Store tokens in local
.envfiles or your secret manager. - Rotate the token immediately if it is exposed.