SCOOP'D Codebase

Codebase for SCOOP'D: Learning Mixed-Liquid-Solid Scooping via Sim2Real Generative Policy.

[Project Page] [arXiv]

SCOOP'D studies robotic scooping in mixed liquid-solid environments. The system collects scooping demonstrations in simulation, learns generative policies with Diffusion Policy, and deploys the learned policy to real-world robot scooping tasks.

This repository is under active cleanup. Full installation instructions, checkpoints, datasets, and training scripts will be released later.

SCOOPD_Codebase/
├── simulator/      # OmniGibson simulation and data collection
├── models/         # Policy and perception models
├── inference/      # Real-world inference and robot execution
└── README.md

• simulator/: simulation environment and heuristic data collection.
• models/diffusion_policy/: Diffusion Policy dependency.
• models/GroudingDINO/: GroundingDINO dependency.
• models/sam2/: SAM2 dependency.
• models/pointnet/: PointNet++ object state regression.
• inference/: RGB-D perception, target localization, policy inference, and robot execution.

Simulation data collection:

python simulator/omnigibson/examples/learning/collect_data_row.py

Real-world inference:

python inference/main_policy_axis_record.py

Before running, please configure local paths, checkpoints, camera topics, robot SDK, robot IP, and calibration files.

@article{wang2025scoop,
  title={SCOOP'D: Learning Mixed-Liquid-Solid Scooping via Sim2Real Generative Policy},
  author={Wang, Kuanning and Gu, Yongchong and Fu, Yuqian and Shangguan, Zeyu and He, Sicheng and Xue, Xiangyang and Fu, Yanwei and Seita, Daniel},
  journal={arXiv preprint arXiv:2510.11566},
  year={2025}
}