We study how generative behavioral-cloning policies represent multimodal expert behavior. In latent-variable policies, posterior-prior regularization can erase action-mode information; preserving it shifts the challenge to deployment-time prior coverage. In action-space policies, smoothness limits the ability to represent many well-separated modes without sharp transitions or off-support bridges. Synthetic multimodal navigation and physical-robot bimodal manipulation experiments support these mechanisms, while deterministic regression remains competitive on standard robotic simulation benchmarks with limited conditional multimodality.
@misc{mazza2026understandingmultimodalitygenerativebehavioral,title={Understanding Multimodality in Generative Behavioral Cloning},author={Mazza, Lorenzo and Datres, Massimiliano and Rodriguez, Ariel and Bodenstedt, Sebastian and Kutyniok, Gitta and Speidel, Stefanie},year={2026},eprint={2605.22493},archiveprefix={arXiv},primaryclass={cs.LG},url={https://arxiv.org/abs/2605.22493},}
Supervised Mixture-of-Experts for Surgical Grasping and Retraction
Lorenzo Mazza, Ariel Rodriguez, Rayan Younis, and 6 more authors
In Robotics: Science and Systems, co-main author , 2026
A supervised Mixture-of-Experts architecture for imitation learning in surgical grasping and retraction. The approach transfers zero-shot to ex vivo tissue, with preliminary in vivo porcine evaluation.
@inproceedings{mazza2026supervised,title={Supervised Mixture-of-Experts for Surgical Grasping and Retraction},author={Mazza, Lorenzo and Rodriguez, Ariel and Younis, Rayan and Lelis, Martin and Hellig, Ortrun and Li, Chenpan and Bodenstedt, Sebastian and Wagner, Martin and Speidel, Stefanie},booktitle={Robotics: Science and Systems},year={2026},}
An Open-Source Robotics Research Platform for Autonomous Laparoscopic Surgery
Ariel Rodriguez, Lorenzo Mazza, Martin Lelis, and 4 more authors
In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), accepted, co-main author , 2026
An open-source, robot-agnostic Remote-Center-of-Motion controller integrated into a ROS platform for teleoperation, demonstration recording, and autonomous laparoscopic policy deployment.
@inproceedings{rodriguez2026platform,title={An Open-Source Robotics Research Platform for Autonomous Laparoscopic Surgery},author={Rodriguez, Ariel and Mazza, Lorenzo and Lelis, Martin and Younis, Rayan and Bodenstedt, Sebastian and Wagner, Martin and Speidel, Stefanie},booktitle={IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},year={2026},}
Open-H-Embodiment: A Large-Scale Dataset for Enabling Foundation Models in Medical Robotics
An open, multi-institution dataset of synchronized medical robotic video and kinematics for training and evaluating robot learning models. I contributed to the data collection pipeline and acquisition of 200 in vivo surgical trajectories.
@article{openh2026embodiment,title={Open-H-Embodiment: A Large-Scale Dataset for Enabling Foundation Models in Medical Robotics},author={{Open-H-Embodiment Consortium}},journal={arXiv preprint arXiv:2604.21017},year={2026},}
LAR-MoE: Latent-Aligned Routing for Mixture of Experts in Robotic Imitation Learning
Ariel Rodriguez, Chenpan Li, Lorenzo Mazza, and 5 more authors
In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), accepted, coauthor , 2026
A two-stage imitation-learning framework that discovers skills without phase labels and uses the learned latent structure to guide Mixture-of-Experts routing. Evaluation on LIBERO and surgical grasping and retraction shows structured expert specialization and zero-shot transfer to ex vivo porcine tissue.
@misc{rodriguez2026larmoelatentalignedroutingmixture,title={LAR-MoE: Latent-Aligned Routing for Mixture of Experts in Robotic Imitation Learning},author={Rodriguez, Ariel and Li, Chenpan and Mazza, Lorenzo and Younis, Rayan and Hellig, Ortrun and Bodenstedt, Sebastian and Wagner, Martin and Speidel, Stefanie},year={2026},eprint={2603.08476},archiveprefix={arXiv},primaryclass={cs.RO},url={https://arxiv.org/abs/2603.08476},}
Workshop papers
Sensorium: An Open-Source Multimodal Data Collection and Curation Library for Embodied Surgical AI Accepted · GenAI4Health workshop, NeurIPS 2026
An open-source Python library for recording, processing, annotating, and curating multimodal surgical data with ROS and ROS 2. Sensorium connects configurable robot and sensor streams to episodic datasets, with evaluation across phantom, ex vivo, and in vivo porcine experiments.
ANCHOR: Safe Diffusion Planning via Analytic Approximation of Doob’s h-function Coauthor · Accepted at STODY, NeurIPS 2026
Safe trajectory planning as conditional generation: an analytic approximation to Doob guidance combines Tweedie posterior moments with signed-distance geometry, guiding a pretrained diffusion model without obstacle-aware retraining. The work establishes safety under explicit regularity assumptions and evaluates distribution recovery and planning on synthetic and Maze2D tasks.
Under review
ANCHOR: Safe Diffusion Planning via Analytic Approximation of Doob’s h-function Co-main author · Conference paper · Under review
Safe diffusion planning, with evaluation across navigation, locomotion, and manipulation.