A4D-Checkpoint

A4D Checkpoint

Fine-tuned CLIP checkpoint for A4D: Functional Latent Spaces for Affordance Reasoning (CoRL 2026).

• Project page: https://a4dance-reasoning.github.io/
• Paper (arXiv): https://arxiv.org/abs/2606.05533
• Dataset: https://huggingface.co/datasets/rohansiva/A4D-dataset

Description

This checkpoint is a CLIP ViT-B/32 model (OpenAI pretrained weights as the base) fine-tuned for affordance classification: given an image and an affordance (e.g. "movable") with its antonym (e.g. "fixed"), the model's image/text embeddings are used to infer whether the affordance applies, by comparing similarity to the affordance term versus its antonym.

Format

The checkpoint is a PyTorch .pt file containing a dict with a model_state_dict key, loadable into an open_clip ViT-B-32 model:

import torch
import open_clip

model, _, preprocess = open_clip.create_model_and_transforms("ViT-B-32", pretrained="openai")
ckpt = torch.load("A4D_model.pt", map_location="cpu")
state_dict = ckpt["model_state_dict"] if "model_state_dict" in ckpt else ckpt
model.load_state_dict(state_dict)
model.eval()

tokenizer = open_clip.get_tokenizer("ViT-B-32")

See the demo notebook (classification_uncertainty.ipynb) in the A4D code repo for a full classification + calibrated-uncertainty walkthrough using this checkpoint.

Citation

@inproceedings{siva2026objectsenablearefunctional,
            title={What Objects Enable, Not What They Are: Functional Latent Spaces for Affordance Reasoning},
            author={Rohan Siva and Neel P. Bhatt and Yunhao Yang and Seoyoung Lee and Nishant Gadde and Christian Ellis and Alvaro Velasquez and Zhangyang Wang and Ufuk Topcu},
            year={2026},
            booktitle={Proceedings of the Tenth Conference on Robot Learning},
            address={Austin, TX, USA},
            publisher={PMLR},
      }