mlsdpk/ConstraintIK

Exploring the Intrinsic Geometry of Diffusion Models with Constrained Inverse Kinematics

Models and data for the CoRL 2026 paper by Miguel Angel Rogel Garcia*, Phone Thiha Kyaw* and Jonathan Kelly (*equal contribution). An earlier version appeared at the RSS 2026 Workshop on Diffusion for Robot Learning.

Paper · Project page · Code

Files

Path	Contents
checkpoints/ur5.pt	UR5 model used throughout the paper
checkpoints/franka.pt	Franka model used throughout the paper
checkpoints/size_sweep/{robot}/{constraint}/n040000.pt	14 single-family models trained on N = 40,000 configurations (training-set size ablation, App. C)
data/ground_truth_ur5.pt, data/ground_truth_franka.pt	Evaluation targets, 50 per constraint family, each with its set of IK solutions
data/train_ur5.pt, data/train_franka.pt	Training sets of the two paper models

Each paper model is a conditional DDPM (25.6M parameters, sigmoid schedule, T = 100) trained for 500k steps on all seven constraint families of one robot. The single-family models use a cosine schedule and 100k steps. A checkpoint stores the model and EMA weights with the normalizers of its training set.

Usage

git clone https://github.com/utiasSTARS/ConstraintIK.git && cd ConstraintIK
pip install -e .
python scripts/download.py                 # paper models and evaluation targets
python scripts/download.py --training      # also the training sets
python scripts/download.py --size_sweep    # also the 14 single-family models

from constraint_ik.checkpoint import load_checkpoint
ik = load_checkpoint("checkpoints/franka.pt", "cuda")

Citation

@inproceedings{rogelgarcia2026intrinsic,
  title     = {Exploring the Intrinsic Geometry of Diffusion Models with Constrained Inverse Kinematics},
  author    = {Rogel Garcia, Miguel Angel and Kyaw, Phone Thiha and Kelly, Jonathan},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2606.26408}
}