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..
Understanding while Exploring: Semantics-driven Active Mapping
Authors: Liyan Chen, Huangying Zhan, Hairong Yin, Yi Xu, Philippos Mordohai
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on the Replica and Matterport3D datasets highlight the effectiveness of Active SGM in active semantic mapping tasks. 4 Experiments and Results |
| Researcher Affiliation | Collaboration | Liyan Chen 1, Huangying Zhan 2, Hairong Yin 3, Yi Xu 2, Philippos Mordohai 1 1 Stevens Institute of Technology 2 Goertek Alpha Labs 3 Purdue University |
| Pseudocode | No | The paper describes the methodology in Section 3.1 and 3.2 using prose and mathematical formulations. No explicit pseudocode or algorithm blocks are provided. |
| Open Source Code | Yes | Our code is avaiable at https://github.com/lly00412/Active SGM.git. |
| Open Datasets | Yes | We evaluate on three photorealistic datasets: Replica [102], Replica SLAM, and MP3D [103]. 1. Replica Dataset [102] URL: https://github.com/facebookresearch/Replica-Dataset License: Research or Education only. 2. Matterport3D Dataset [103] URL: https://niessner.github.io/Matterport/ License: Non-commercial |
| Dataset Splits | Yes | To improve semantic prediction accuracy, we collect 500 RGBSemantic frames from each scene and fine-tune One Former separately on Replica and MP3D, training for 3,000 steps per scene. The novel trajectories described in Table 1 of the main paper are used as the test set. These trajectories are distinct from those used for fine-tuning. The train/test Top-1 accuracy is reported in Table S.1. |
| Hardware Specification | Yes | All experiments were conducted on two NVIDIA RTX A6000 GPUs. We conducted the experiments on a server with 2 NVIDIA RTX A6000 GPUs and an Intel i9-10900X CPU with 20 cores. |
| Software Dependencies | Yes | Our Active SGM is implemented with python 3.8 and CUDA 11.7. We use Habitat [101] to generate RGB-D frames and One Former [17] for semantic segmentation. ... We adopt the streamlined approach proposed in Spla TAM [51]. ... We also implement SGS-SLAM [7] for comparative analysis. |
| Experiment Setup | Yes | We set λHD = 0.8 and λcos = 0.2 to balance their contributions. The Exploration Map uses a voxel size of 5 cm. Each experiment runs for 2,000 steps on Replica and 5,000 on MP3D, with early termination if the exploration candidate pool is exhausted. ... coarse stage samples on a single height plane with larger steps (v1 = 1) and fewer directions (v2 = 5); the fine stage increases density with smaller steps (v1 = 0.5), multiple heights, and more directions (v2 = 15). |