[CVPR 2026] StaCOM: Stability-Driven Motion Generation for Object-Guided Human-Human Co-Manipulation

Official codebase for CVPR 2026 paper "Stability-Driven Motion Generation for Object-Guided Human-Human Co-Manipulation".

Jiahao Xu, Xiaohan Yuan, Xingchen Wu, Chongyang Xu, Kun Li, Buzhen Huang
Tianjin University, National University of Singapore, Sichuan University

Project · Paper

---

Installation

The code is tested on Ubuntu with a single RTX 4090 GPU (24GB).

Create the conda environment:

conda create -n stacom python=3.10
conda activate stacom

Install PyTorch with CUDA 11.8:

pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 torchaudio==2.1.2+cu118 --index-url https://download.pytorch.org/whl/cu118

Install other dependencies:

pip install -r requirements.txt

Download the official SMPL-X model from the SMPL-X website and place it in data/smplx/.

Demo

Demo assets (checkpoints + SMPL-X neutral model) are provided here:

• Google Drive: https://drive.google.com/drive/folders/17oLiCvTHiHTnGxfUHlmu687rCeGwu4Rk?usp=drive_link

Download and place the files as follows:

• contact_epoch200.pkl -> output/contact_epoch200.pkl
• hoi_epoch200.pkl -> output/hoi_epoch200.pkl
• SMPLX_NEUTRAL.pkl -> data/SMPLX_NEUTRAL.pkl

Recommended folder structure:

StaCOM/
├── data/
│   └── SMPLX_NEUTRAL.pkl
└── output/
    ├── contact_epoch200.pkl
    └── hoi_epoch200.pkl

Input format

Three files are required:

• object.obj — object mesh in its local coordinate frame
• trajectory.npy — object 6D pose sequence, shape (T, 4, 4)
• affordance.npz — per-point affordance scores, must contain key sampled_scores of shape (N,)

Run

Run the motion generation demo with a trained checkpoint:

xvfb-run -a -s "-screen 0 1024x768x24" python demo.py \
    --obj-mesh     data/test/01/box001.obj \
    --obj-traj     data/test/01/trajectory.npy \
    --affordance   data/test/01/affordance.npz \
    --contact-ckpt output/contact_epoch200.pkl \
    --motion-ckpt  output/hoi_epoch200.pkl \
    --body-model   data/SMPLX_NEUTRAL.pkl \
    --output-dir   output/

xvfb-run starts a virtual X display for headless/offscreen rendering on servers without a desktop session (common for remote Linux machines).

Install it with:

# Ubuntu / Debian
sudo apt-get update && sudo apt-get install -y xvfb

The output video is saved to output/res_20260325_143022.mp4.

Optional arguments:

Argument	Default	Description
--gpu-index	0	CUDA device index
--physics	off	Enable stability-driven physics simulation (CMA-ES).

Run the visualization demo below for contact point:

python vis_contact.py

The demo expects uploaded inputs such as:

• mesh (.obj)
• object trajectory (trajectory.npy)
• affordance (affordance.npz)
• GT contact (gt_contact.npz)

---

Condition Generation

Generate necessary condition data with:

python utils/data_collection.py --config=cfg_files/config.yaml

BPS	Affordance	Object Trajectory + Contact Points
Basis Point Set representation encoding the object geometry and distance features.	Predicted interaction affordance regions on the object surface.	Object motion trajectory together with human-object contact locations.

Training

SDF loss is required for penetration evaluation.

Download the dataset from:

(To be released)

Place the dataset under data/ as specified by --data_folder, then run the following to train the motion generation model:

python main.py \
    --mode        train \
    --data_folder data \
    --trainset    "CORE4D_real CORE4D_syn" \
    --testset     CORE4D_S1 \
    --model       interhuman_flow_BPS_prior \
    --epoch       2000 \
    --batchsize   4 \
    --lr          0.0001 \
    --worker      6 \
    --output      output

Citation

@inproceedings{xu2026stability,
  title={Stability-Driven Motion Generation for Object-Guided Human-Human Co-Manipulation},
  author={Xu, Jiahao and Yuan, Xiaohan and Wu, Xingchen and Xu, Chongyang and Li, Kun and Huang, Buzhen},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}