GNN-Transformer Task Planning Enhanced with Semantic-Driven Data Augmentation
Authors: Soojin Jeong, Seongwan Byeon, Sangwoo Kim, HyeokJun Kwon, Yoonseon Oh
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our model against a comprehensive set of baselines. Each leverages distinct methodologies within the realm of LLM and planning... The results of compared baselines in the VH environment are shown in Table 1. Our model achieved an 18.1% higher success rate than the best-performing baseline in seen environments. |
| Researcher Affiliation | Academia | Soojin Jeong*, Seongwan Byeon*, Sangwoo Kim, Hyeok Jun Kwon, Yoonseon Oh Department of Electronic Engineering, Hanyang University, Seoul, Korea EMAIL |
| Pseudocode | No | The paper describes the proposed methods, GNN-Transformer Task Planning and Semantic-Driven Data Augmentation, in detailed text descriptions and illustrates them with figures, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions using 'ROS2 humble environment with open-source packages' for the robot control system, and specifically 'Moveit2 and NAV2 with Slam Toolbox' for manipulation and navigation. However, it does not provide an explicit statement or link for the source code of the methodology described in this paper. |
| Open Datasets | No | The paper describes generating a dataset of '14k instruction-annotated tasks in the Virtual Home environment' using an automated data augmentation method. While it mentions using the Virtual Home (Puig et al. 2018) simulator and collecting seed tasks from other cited works (Singh et al. 2023), there is no explicit statement, link, or citation indicating that the specific 14k dataset generated by the authors is publicly available. |
| Dataset Splits | Yes | We split the proposed dataset into four subsets: a training set, a validation set, a seen environment test set, and unseen environment test set... we distributed the remaining data into training, validation, and seen environment test sets in a 0.7:0.15:0.15 ratio, ensuring no overlap of environment and instruction pairs between training and test sets. The dataset includes 10,054 for training, 1,918 for validation, 2,044 for seen tests, and 650 for unseen tests. |
| Hardware Specification | No | The paper mentions using 'a mobile manipulator combining a Husky mobile base robot with a Franka Emika Panda arm' for real-world experiments. However, it does not provide specific details on the computational hardware (e.g., GPU models, CPU models, or memory specifications) used for training the models or running simulations. |
| Software Dependencies | Yes | The robot control system is implemented on ROS2 humble environment with open-source packages. Moveit2 and NAV2 with Slam Toolbox are utilized for manipulation and navigation, respectively. |
| Experiment Setup | Yes | In this paper, the number of GNN layers NGNN is set as 3 and the dimensions of the updated node, edge, and instruction feature are reduced to 256, 128, 64, respectively. ... During training, we introduce the node dropout technique... We randomly drop out 30%, 60%, and 90% of Ob... For training, we employ Cross-entropy as the loss function in our model as follows: Ltotal = αLaction + βLobject... |