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..
TP-MDDN: Task-Preferenced Multi-Demand-Driven Navigation with Autonomous Decision-Making
Authors: Shanshan Li, Da Huang, Yu He, Yanwei Fu, Yu-Gang Jiang, Xiangyang Xue
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments demonstrate that our approach outperforms state-of-the-art baselines in both perception accuracy and navigation robustness. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art baselines, with ablation studies confirming the effectiveness of each component. |
| Researcher Affiliation | Collaboration | Shanshan Li1, Da Huang23, Yu He13, Yanwei Fu13 , Yu-Gang Jiang1, Xiangyang Xue1 1Fudan University 2Shanghai Jiao Tong University 3Shanghai Innovation Institution |
| Pseudocode | No | The paper describes the system architecture and modules using text and figures (Figure 2, Figure 3), but it does not contain any explicitly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The code declaration will be open source. |
| Open Datasets | Yes | We use AI2-THOR [69] as our simulator and Proc Thor as our scene dataset [66]. From the test set of Proc THOR [66], we generated 200 long-horizon instructions, each containing three subtasks, spanning 68 rooms. |
| Dataset Splits | Yes | From the test set of Proc THOR [66], we generated 200 long-horizon instructions, each containing three subtasks, spanning 68 rooms. We used each of Deep Seek-V3 [9] and GPT-4o [10] to generate 100 task-preferenced, multi-demanddriven unseen instructions in test scenarios (totaling 200 commands). |
| Hardware Specification | Yes | All experiments can be run on a single NVIDIA H100 80GB GPU. |
| Software Dependencies | No | The paper mentions specific models like Ram-Grounded-SAM model [67, 68], Deep Seek V3 [9], GPT-4o [10], and Qwen2-5-VL-72B [71, 72], and simulators like AI2-THOR [69], but it does not provide specific version numbers for software libraries, programming languages, or other ancillary software components. |
| Experiment Setup | Yes | In all experimental settings, the success distance threshold ฯตdis is 1.5 meters, the maximum step count Lenmax is 50, and the tolerance for repeated failed attempts on the same object ntolerance within Extrainfo is 2. The obstacle avoidance distance ฯobs used during affordance map computation is 0.25 meters. When processing input data, we set the camera resolution to 300ห300 and the horizontal field of view (HFo V) to 90 for the agent. |