• Overview
• Installation
• Data
• Training
• Evaluation
• Checkpoints
• Acknowledgement
• Citation
• License

This repository is the implementation code of the NeurIPS 2025 paper: Hyper-GoalNet: Goal-Conditioned Manipulation Policy Learning with HyperNetworks.

In this repo, we provide our implementation code of training and evaluation in simulation. To simplify the process of reproducing our results, we also provide the datasets generated from MimicGen and the corresponding checkpoints for each task.

1. Create and activate conda environment

   conda create -n hyperpolicy python=3.8
   conda activate hyperpolicy
   

2. Clone the repository

   git clone https://github.com/wantingyao/hyper-goalnet.git
   

3. Install dependency packages

   python -m pip install pip==23.3.1
   cd hyper-goalnet
   pip install -r requirements.txt
   

4. We use r3m pre-trained weights for our visual encoder.

   git clone https://github.com/facebookresearch/r3m.git
   cd r3m
   pip install -e .
   

5. The simulation result of Hyper-GoalNet is tested on robosuite benchmark using datasets generated by mimicgen. Please follow the steps below to install the required dependencies.

   # Install robosuite
   cd ..
   git clone https://github.com/ARISE-Initiative/robosuite.git
   cd robosuite
   git checkout b9d8d3de5e3dfd1724f4a0e6555246c460407daa
   pip install -e .
   cd ..
   git clone https://github.com/ARISE-Initiative/robosuite-task-zoo
   cd robosuite-task-zoo
   git checkout 74eab7f88214c21ca1ae8617c2b2f8d19718a9ed
   pip install -e .
   # Install mimicgen
   cd ..
   git clone https://github.com/NVlabs/mimicgen.git
   cd mimicgen
   pip install -e .
   

6. We modified robosuite for our evaluation on randomly generated testing data from mimicgen. After installing robosuite, please copy the following codes to the end of the file robosuite/robosuite/environments/base.py.

       def save_state_dict(self):
           return self.sim.get_state().flatten()
       
       def set_specified_state(self, specified = None, value = None):
           if(specified):
               self.sim.set_state_from_flattened(value)
               self.sim.forward()
               
       def get_obs(self):
           observations = (
               self.viewer._get_observations(force_update=True)
               if self.viewer_get_obs
               else self._get_observations(force_update=True)
           )
           # Return new observations
           return observations

The training and testing datasets used in our experiments are available for download from link. Note that these datasets were generated by mimicgen, you can also follow the mimicgen documentation to generate your own custom datasets.

Ensure your data is placed in the correct folders as specified in the project structure.

cd ../hypergoal
mkdir -p datasets/{testing_data,training_data}

.
├── hyper-goalnet/
│   ├── algo/
│   ├── configs/
│   ├── datasets/
│   │   ├── testing_data/
│   │   │   └── ${task}.hdf5 	# place your testing data here
│   │   └── training_data/
│   │       └── ${task}.hdf5 	# place your training data here
│   ├── models/
│   ├── scripts/
│   ├── utils/
├── mimicgen/
├── r3m/
├── robosuite/
├── robosuite-task-zoo/

Run the following script to start training. The training configuration is located at hypergoal/configs/config.yaml. The default training task is coffee_d0.

python scripts/train.py

Update hypergoal/configs/eval_config.yaml with your checkpoint path, then run the evaluation script.

python scripts/eval.py

If you have a remote X server, set the DISPLAY variable to your actual display number:

# Check current display
echo $DISPLAY
# Set display variable (replace :0 with your actual display number)
export DISPLAY=:0

You can download our trained checkpoints from Link.

Please cite Hyper-GoalNet if you find this repository helpful:

@article{zhou2025hyper,
  title={Hyper-GoalNet: Goal-Conditioned Manipulation Policy Learning with HyperNetworks},
  author={Zhou, Pei and Yao, Wanting and Luo, Qian and Zhou, Xunzhe and Yang, Yanchao},
  journal={arXiv preprint arXiv:2512.00085},
  year={2025}
}

Licensed under the MIT License.