Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

HOI-Dyn: Learning Interaction Dynamics for Human-Object Motion Diffusion

Authors: Lin Wu, Zhixiang Chen, Jianglin Lan

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive qualitative and quantitative experiments, we demonstrate that our approach not only enhances the quality of HOI generation but also establishes a feasible metric for evaluating the quality of generated interactions.
Researcher Affiliation Academia Lin Wu James Watt School of Engineering University of Glasgow EMAIL Zhixiang Chen School of Computer Science University of Sheffield EMAIL Jianglin Lan James Watt School of Engineering University of Glasgow EMAIL
Pseudocode No The paper describes the methodology using mathematical equations and textual descriptions, for example in sections '3.1 Interaction Dynamics' and '3.2 Conditional Motion Diffusion', but no specific block explicitly labeled 'Pseudocode' or 'Algorithm' is provided.
Open Source Code Yes Project website:https://wulin97.github.io/hoi-dyn Our project website is available at https://wulin97.github.io/hoi-dyn/, and the code is available at https://github.com/AIR-Lan/HOI-Dyn.
Open Datasets Yes We train and evaluate HOI generation using two datasets: (i) Full Body Manipulation, which provides 10 hours of high-quality paired object and human motion data involving 15 different objects [1]. We apply the OMOMO partitioning to this dataset: 15 subjects for training and 2 for testing; and (ii) 3D-FUTURE, which consists of 3D models of various furniture items [30].
Dataset Splits Yes We apply the OMOMO partitioning to this dataset: 15 subjects for training and 2 for testing
Hardware Specification Yes All experiments are conducted on a single NVIDIA RTX A4500 GPU, with total training time around 10 hours.
Software Dependencies No The paper mentions using the Adam optimizer [32] and Cosine Annealing Warm Restarts [33] but does not provide specific version numbers for these or any other software libraries or programming languages used.
Experiment Setup Yes We train the interaction dynamics using Full Body Manipulation with K = 2 as the maximum prediction horizon. The network has 0.5M parameters. We use the Adam optimizer [32] with a learning rate of 1 10 3, adjusted via Cosine Annealing Warm Restarts [33] for 150 epochs with a batch size of 32. The model is then transferred to the HOI motion diffusion task, where fine-grained driver-responder relationships are encouraged during denoising, as in (10). Training is done from scratch with a learning rate of 1 10 4, batch size 32, and 100,000 steps.