EPFL-VILAB

/

TST-ProcTHOR-adapted

like 0

Follow

EPFL VILAB 94

Hi EPFL-VILAB team and community 👋

I found this model: EPFL-VILAB/TST-ProcTHOR-adapted. It looks like a vision model trained/adapted for the ProcTHOR environment, but I can’t find a README or usage instructions on the repo page.

Could someone share the recommended way to run it (inference and/or evaluation)?

What I’m trying to do:

• Load the model from Hugging Face
• Run inference on a sample ProcTHOR observation (e.g., RGB frame / optional depth / optional segmentation)
• (If applicable) run the model inside a ProcTHOR loop for evaluation

A few concrete questions:

1. What is the expected input format? (RGB only? RGB+depth? tokenized prompts? normalization / resolution?)

2. What is the output? (embeddings, action logits, goal-conditioned predictions, captions, etc.)

3. Which codebase should be used with this checkpoint?

   • Is there a reference repo (GitHub) for TST + ProcTHOR training/inference?
   • Any exact commit/version or branch that matches this checkpoint?

4. Any required dependencies and versions? (PyTorch version, transformers, habitat/ai2thor/procthor packages, etc.)

5. Is there an example config (YAML/JSON) or a minimal script to reproduce a forward pass?

If someone can provide either:

• a minimal “load + forward pass” snippet, or
• a link to the matching repo / docs / config,

…I’d really appreciate it. I can also test and report back with the exact steps once I get guidance.

Thanks!