metadata
license: apache-2.0
tags:
- robotic manipulation
Cortical Policy: A Dual-Stream View Transformer for Robotic Manipulation
Official pretrained weights for Cortical Policy (ICLR 2026). This model integrates static and dynamic visual streams to enhance spatial reasoning and dynamic adaptation for robotic manipulation.
📂 Repository Structure
This repository contains the necessary checkpoints to run and evaluate the Cortical Policy framework. The weights are organized into two main categories:
dynamic_view_stream_backbone_weights/: Contains the pretrained weights for the Gaze Model (dynamic-view stream), pre-trained on Ego4D and fine-tuned for position awareness.policy_weights/: Contains the final trained policy checkpoints (model.pth) along with the corresponding YAML configuration files (exp_cfg.yaml,mvt_cfg.yaml).
🛠️ Usage
Please refer to our official GitHub repository for the complete installation guide, data preparation steps, and training/evaluation scripts:
👉 iLearn-Lab/ICLR26-Cortical_Policy
📝 Citation
If you find our model or code useful in your research, please cite our ICLR 2026 paper:
@inproceedings{zhang2026cortical,
title={Cortical Policy: A Dual-Stream View Transformer for Robotic Manipulation},
author={Xuening Zhang and Qi Lv and Xiang Deng and Miao Zhang and Xingbo Liu and Liqiang Nie},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026}
}