PoseShield: Neural Collision Fields for Human Self-Collision Resolution (ECCV 2026)

Zhengyuan Li¹  Zeyun Deng¹  Yifan Shen³  Liangyan Gui³  Miaolan Xie¹ 
Joseph Campbell¹  Xifeng Gao²  Kui Wu²  Zherong Pan²  Aniket Bera<sup>1 

1</sup>Purdue University  ²LightSpeed Studios  ³University of Illinois Urbana-Champaign

PoseShield is a post-hoc self-collision resolver for SMPL-H poses and human motion sequences. It uses a learned neural collision field as a differentiable constraint, so existing poses and motions can be repaired as a post-processing step without even knowing the upstream system that produced them.

• Quick pose and motion demos with released checkpoints.
• Humans with Collisions (HwC) data for training, evaluation, and the 500-pose benchmark.
• Exact-FCL validation plus recommended Blender visualization and legacy HTML fallback.
• Experimental shape-aware SAField demo that resolves a fixed sample and exports OBJ meshes.

PoseShield treats self-collision correction as a post-hoc optimization problem: given a self-intersecting SMPL-H pose or motion, find a nearby collision-free result while preserving the original pose semantics and motion dynamics. The input can come from a generative model, motion capture cleanup pipeline, dataset preprocessing workflow, or any other source that can be represented in the public SMPL-H/HY-Motion-compatible layouts below.

The core component is a neural collision field defined directly in SMPL-H pose space. The field is trained to be positive for collision-free poses and negative for self-intersecting poses, so it can be used as a differentiable collision constraint. We regularize this field with an Eikonal-style objective, encouraging non-vanishing gradients near the collision boundary and making gradient-based optimization more stable.

At inference time, PoseShield uses this learned field in two ways:

• For single poses, it solves a constrained optimization problem that minimally changes the input SMPL-H body rotations while moving the pose into the collision-free region.
• For motion sequences, it reuses the same learned collision field inside a two-stage latent optimization pipeline: Stage 1 fits the motion model latent to the input sequence, and Stage 2 resolves self-collisions while preserving hand motion, temporal dynamics, and the original global translation.

The main method studied in the paper assumes the neutral SMPL-H body model with betas=None; subject-specific body-shape parameters are not passed to the SMPL-H layer. To apply the method to other SMPL body shapes, character-specific SMPL humans, Momentum Human Rig, or other human parametric models, users may need to build a custom collision dataset for the target body model and retrain the collision field. In this release, we include a preliminary experimental shape-aware collision-field feature as an initial step beyond the paper setting.

conda env create -f environment.yml
conda activate poseshield
pip install -e .

We test the code on Python 3.10 and PyTorch with CUDA. The release environment.yml pins the dependency versions that are most likely to affect reproducibility, including PyTorch, NumPy, python-fcl, transformers, and diffusers.

Register and download the Extended SMPL+H model from the MANO website.

mkdir -p deps/body_models/smplh
cp smplh/neutral/model.npz deps/body_models/smplh/SMPLH_NEUTRAL.npz

PoseShield currently uses the neutral SMPL-H model.

Download and extract the PoseShield external assets at the repository root:

unzip PoseShield_release_dependencies_20260628.zip -d .
unzip PoseShield_release_pose_data_20260628.zip -d .
unzip PoseShield_release_motion_data_20260708.zip -d .
unzip PoseShield_release_safield_demo_20260703.zip -d .  # optional experimental SAField demo

The release asset packages are available from the PoseShield Google Drive folder.

Validate the asset layout with:

python tools/check_assets.py

deps/
+-- body_models/
|   +-- smplh/
|       +-- SMPLH_NEUTRAL.npz
+-- topology_distances_30_60.npz
data/
+-- dataset/                     # Humans with Collisions (HwC) train/test data
|   +-- train_list.csv
|   +-- test_list.csv
|   +-- augmented_data/
|   |   +-- *.npz
|   +-- gt_data/
|       +-- *.npz
+-- dataset_test/                # HwC 500-pose collision-resolution benchmark
|   +-- *.pkl
|   +-- *.obj
|   +-- *.png
+-- motion_canonical/            # 100 canonical MotionFix motion sequences
    +-- motionfix_*_135.npy
ckpts/
+-- poseshield/
|   +-- config.yaml
|   +-- config_elu.yaml
|   +-- model.pth
|   +-- model_elu.pth
+-- tencent/
    +-- HY-Motion-1.0-Lite/
        +-- latest.ckpt
        +-- config.yaml  # or config.yml
        +-- stats/
            +-- Mean.npy
            +-- Std.npy
experimental/
+-- safield_demo/                # optional experimental shape-aware demo
    +-- sa_model.pth
    +-- sa_config.yaml

Small demo inputs are provided in demo_asset/.

python demos/demo_pose.py

Outputs are written to demos/output/.

Run the interactive demo:

bash demos/demo_motion.sh

SAMPLE=motion_sample2.npy
STEM=${SAMPLE%.npy}

python -m poseshield.hymotion.dno.run_dno_stage1 \
    --model_path ckpts/tencent/HY-Motion-1.0-Lite \
    --motion_file demo_asset/$SAMPLE \
    --output_dir demos/output_motion/$STEM

python -m poseshield.hymotion.dno.run_dno_stage2 \
    --model_path ckpts/tencent/HY-Motion-1.0-Lite \
    --motion_file demo_asset/$SAMPLE \
    --stage1_z demos/output_motion/$STEM/stage1_z.pt \
    --output_dir demos/output_motion/${STEM}_stage2

optimized_motion.npy
optimized_z.pt
summary.json
args.json

shape: [frames, 135]

[0:132]   22 joints × 6D rotations, HY-Motion column-interleaved layout
[132:135] absolute global translation [x, y_up, z_forward]

Y-up
X = right
Y = height/up
Z = forward
frame 0 human facing +Z

python tools/evaluate_exact_fcl.py \
    --motion demos/output_motion/${STEM}_stage2/optimized_motion.npy \
    --output-dir demos/output_motion/${STEM}_stage2/exact_fcl

bash demos/demo_blender_contact_render.sh

python tools/generate_motion_html.py \
    --sequence $STEM \
    --original demo_asset/$SAMPLE \
    --optimized demos/output_motion/${STEM}_stage2/optimized_motion.npy \
    --output-dir demos/output_motion/${STEM}_stage2/visualization

python poseshield/pose/resolve_dataset_test_slsqp.py \
    --config-path ckpts/poseshield/config.yaml \
    --model-path ckpts/poseshield/model.pth \
    --n-samples 500 \
    --cost-type weighted \
    --threshold 0.1 \
    --max-itr 300 \
    --save

Pose-level SLSQP logs report three separate statuses:

• solver_success: whether SciPy SLSQP terminated with its success flag.
• constraint_satisfied: whether the learned collision-field constraint meets the requested threshold.
• exact_collision_free: whether the final SMPL-H mesh is collision-free under exact-FCL validation.

The default SLSQP iteration budget is --max-itr 300.

python poseshield/pose/evaluate.py \
    --config-path ckpts/poseshield/config.yaml \
    --model-path ckpts/poseshield/model.pth

Train the collision field from scratch:

python -m poseshield.pose.train --config-path config_files/basic_config.yaml

Checkpoints and logs are saved to experiments/<EXP_NAME>/.

We also include a minimal standalone SAField demo that conditions the collision field on SMPL body-shape coefficients. Starting from a fixed colliding pose, the demo loads the released experimental checkpoint, resolves the pose for two body shapes, and can export the input and resolved meshes as OBJ files.

This component is provided as an experimental extension rather than the primary PoseShield release path. See experimental/safield_demo/ for the model config, fixed sample, and command to reproduce the OBJ outputs.

If you find our work useful in your research, please consider citing:

@article{li2026poseshield,
  title={PoseShield: Neural Collision Fields for Human Self-Collision Resolution},
  author={Li, Zhengyuan and Deng, Zeyun and Shen, Yifan and Gui, Liangyan and Xie, Miaolan and Campbell, Joseph and Gao, Xifeng and Wu, Kui and Pan, Zherong and Bera, Aniket},
  journal={arXiv preprint arXiv:2606.29686},
  year={2026}
}

This project builds upon SMPL-X, HY-Motion-1.0, python-fcl, Diffusion-Noise-Optimization, and the MotionFix dataset.

This project is licensed under the MIT License. External body models, datasets, and upstream model checkpoints may be subject to their own licenses.