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..
Learning Spatial-Aware Manipulation Ordering
Authors: Yuxiang Yan, Zhiyuan Zhou, Xin Gao, Guanghao Li, Shenglin Li, Jiaqi Chen, Qunyan Pu, Jian Pu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Order Mind on our Manipulation Ordering Benchmark, comprising 163,222 samples of varying difficulty. Extensive experiments in both simulation and real-world environments demonstrate that our method significantly outperforms prior approaches in effectiveness and efficiency, enabling robust manipulation in cluttered scenes. |
| Researcher Affiliation | Collaboration | 1 Fudan University 2 Shanghai Yin Cheng Intelligent CO., LTD 3 Stanford University |
| Pseudocode | No | The paper describes methods in prose, but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | We provide reproducible code along with detailed instructions, ensuring that the main experimental results can be faithfully reproduced. |
| Open Datasets | Yes | For simulation, we utilized Pybullet [43] as the engine and designed cluttered scenes using the YCB object set [44]. |
| Dataset Splits | Yes | The simulation dataset consists of 161,722 RGB-D images for training, 1,500 images for validation. Closed-loop testings are performed within the Py Bullet simulator. For real-world experiments, we collected 26,324 images for training and 6,581 for validation. |
| Hardware Specification | Yes | We trained our model on a single RTX 4090 GPU using a batch size of 24. |
| Software Dependencies | No | The paper mentions software like Pybullet and AdamW optimizer, but does not provide specific version numbers for these or other key software dependencies. |
| Experiment Setup | Yes | We trained our model on a single RTX 4090 GPU using a batch size of 24. The optimizer is Adam W [45] with an initial learning rate of 2 10 4 and a weight decay of 0.01. The learning rate follows a cosine annealing schedule [46]. Training proceeds in two stages. The first stage focuses on pre-training the image backbone [47, 48] for 50 epochs. In the second stage, the full model is trained end-to-end for 10 epochs, enabling the system to learn order-aware manipulation directly from sufficient spatial representation. |