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..
HEIR: Learning Graph-Based Motion Hierarchies
Authors: Cheng Zheng, William Koch, Baiang Li, Felix Heide
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our hierarchical reconstruction approach on three examples: 1D translational motion, 2D rotational motion, and dynamic 3D scene deformation via Gaussian splatting. Experimental results show that our method reconstructs the intrinsic motion hierarchy in 1D and 2D cases, and produces more realistic and interpretable deformations compared to the baseline on dynamic 3D Gaussian splatting scenes. |
| Researcher Affiliation | Collaboration | 1Princeton University 2Torc Robotics chengzh, william.koch, baiang.li, EMAIL |
| Pseudocode | No | The paper describes the proposed method using mathematical formulations (e.g., Equation 1, 3, 4, 5, 7) and textual descriptions of its components (encoder, decoder, training objective, etc.), but it does not include a dedicated, structured pseudocode or algorithm block. |
| Open Source Code | No | Project Page: https://light.princeton.edu/HEIR/ The code and data will be released upon publication. |
| Open Datasets | Yes | We next validate the method on 3D dynamic Gaussian splatting with experiments on D-Ne RF dataset [25], which contains a variety of rigid and non-rigid deformations of various objects. [25] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318 10327, 2021. |
| Dataset Splits | No | The paper describes generating synthetic datasets for 1D motion and a planetary system. For instance, 'We create a minimal point set Xt RN with N = 11 nodes and t {0, . . . , T}, T = 200 frames.' For the D-Ne RF dataset, it mentions evaluating on 'known dynamic poses' and 'dynamic scenes from D-Ne RF dataset', but it does not specify explicit training/test/validation splits for these datasets within the context of the experiments conducted in this paper. |
| Hardware Specification | Yes | The supplemental material specifies the hardware (NVIDIA RTX 3090) and approximate runtimes for different experiments, which are sufficient to estimate required computational resources. |
| Software Dependencies | No | The paper mentions using a 'graph attention layer [33]' and the 'Gumbel-Softmax trick [12]' which refer to specific methods or algorithms. However, it does not provide specific version numbers for any software libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages used in their implementation. |
| Experiment Setup | Yes | We train 1,000 independent models for 1,000 epochs each, annealing the Gumbel Softmax temperature τ from 1.5 to 0.3. We use λ r = 12.0 to penalize deviations in distance, λδ = 0.8 to regularize relative velocity, and λ θ = 0.0 we do not want to penalize relative angular velocity here, on the contrary. The Laplacian connectivity prior is enforced with weight λΛ = 6.0. |