CooHOI: Learning Cooperative Human-Object Interaction with Manipulated Object Dynamics
Authors: Jiawei Gao, Ziqin Wang, Zeqi Xiao, Jingbo Wang, Tai WANG, Jinkun Cao, Xiaolin Hu, Si Liu, Jifeng Dai, Jiangmiao Pang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To validate the effectiveness of our framework, we conducted experiments where we trained control policies for two humanoid characters to carry various long objects, such as boxes and sofas. Our results demonstrate that our framework enables these characters to exhibit natural-looking behaviors while successfully completing cooperative tasks, utilizing only motion capture data from one single agent. We compared our approach against the baseline method of training from scratch and performed detailed ablation studies to evaluate the impact of our design decisions, also testing the limitations of our framework. |
| Researcher Affiliation | Collaboration | Jiawei Gao 1,2, Ziqin Wang 1,3, Zeqi Xiao1,4, Jingbo Wang1, Tai Wang1, Jinkun Cao5, Xiaolin Hu2, Si Liu 3, Jifeng Dai 2, Jiangmiao Pang 1 1Shanghai AI Laboratory, 2Tsinghua University, 3Beihang University, 4Nanyang Technological University, 5Carnegie Mellon University, |
| Pseudocode | No | The paper describes algorithms and formulations using equations but does not provide a pseudocode or algorithm block labeled as such. |
| Open Source Code | Yes | We fully described our method's details in our paper and also provided the code. |
| Open Datasets | Yes | Our primary source of motion data is the AMASS dataset [17], which provides motions encoded in SMPL [22] parameters. |
| Dataset Splits | No | The paper describes general training setups, hyperparameter tuning, and cross-validation is not mentioned. While it mentions '4096 environments' and '4 random seeds' for evaluation, it does not specify explicit dataset splits (e.g., percentages or counts) for training, validation, and testing phases. |
| Hardware Specification | Yes | Our experiments are conducted on the Isaac Gym simulator [18] using a single Nvidia GTX 3090Ti GPU. |
| Software Dependencies | No | The paper mentions several software components like Isaac Gym simulator, PPO, Adam optimizer, but it does not specify version numbers for these or other general software dependencies like Python or PyTorch. |
| Experiment Setup | Yes | Table 3: Hyperparameters for Coo HOI. Parameter Value Number of Environments 4096 w G Task-Reward Weight 0.5 w S Style-Reward Weight 0.5 PPO Minibatch Size 16384 AMP Minibatch Size 4096 Horizon Length 32 Learning Rate 2e 5 PPO Clip Threshold ϵ 0.2 γ Discount 0.99 GAE (λ) 0.95 T Episode Length 600 |