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..
GaussianFusion: Gaussian-Based Multi-Sensor Fusion for End-to-End Autonomous Driving
Authors: Shuai Liu, Quanmin Liang, Zefeng Li, Boyang Li, Kai Huang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on the NAVSIM and Bench2Drive benchmarks demonstrate the effectiveness and robustness of the proposed Gaussian Fusion framework. The source code is released at https://github.com/Say2L/Gaussian Fusion. We evaluate Gaussian Fusion on the planning-oriented NAVSIM dataset [11]. Utilizing the V299 backbone [27], our approach achieves 92.0 PDMS [11], significantly surpassing current stateof-the-art methods. To further evaluate the generalization and robustness of our framework, we conduct experiments on the closed-loop benchmark Bench2Drive [23], where the results consistently demonstrate the effectiveness of Gaussian Fusion. Section 4: Experiments. 4.1 Benchmark and Metric. 4.2 Implementation Details. 4.3 Comparison with State-of-the-Art Methods. 4.4 Ablation Study. 4.5 Qualitative Comparison. |
| Researcher Affiliation | Academia | Shuai Liu, Quanmin Liang, Zefeng Li, Boyang Li , Kai Huang School of Computer Science and Engineering, Sun Yat-sen University {liush376@mail2, liby83@mail, huangk36@mail}.sysu.edu.cn |
| Pseudocode | No | The paper describes the methodology using prose and mathematical equations in sections 3.1, 3.2, and 3.3, but it does not include a clearly labeled pseudocode block or algorithm. |
| Open Source Code | Yes | Extensive experiments on the NAVSIM and Bench2Drive benchmarks demonstrate the effectiveness and robustness of the proposed Gaussian Fusion framework. The source code is released at https://github.com/Say2L/Gaussian Fusion. |
| Open Datasets | Yes | We evaluate models on NAVSIM [11] and Bench2Drive [23] benchmarks. NAVSIM, built upon the Open Scene dataset [10], provides 120 hours of challenging driving data with high-resolution camera images and Li DAR inputs spanning up to 1.5 seconds. Bench2Drive [23], based on the CARLA simulator [12], evaluates E2E autonomous driving across 220 routes that cover 44 interactive scenarios under diverse conditions. |
| Dataset Splits | Yes | For the NAVSIM benchmark, we use the NAVSIM train split for training. We benchmark Gaussian Fusion against leading state-of-the-art (SOTA) methods on the NAVSIM navtest split. We further conduct experiments on the closed-loop benchmark, Bench2Drive, to compare our method with existing SOTA E2E methods. |
| Hardware Specification | Yes | The latency is measured by an RTX3090. |
| Software Dependencies | No | The paper mentions using the Adam W optimizer [34] and backbone networks like ResNet34 [14] and V2-99 [27], but does not specify software dependencies like programming language versions, library versions (e.g., PyTorch, TensorFlow), or CUDA versions. |
| Experiment Setup | Yes | In our main experiments, the number of Gaussians is set to 512, and each Gaussian feature has a dimensionality of 128. We adopt 4 Gaussian Encoder blocks and 2 cascade planning blocks. The number of anchor trajectories is set to 20 following [32]. Training is performed using the Adam W optimizer [34], with 50 epochs, a weight decay of 1 10 4, and a maximum learning rate of 6 10 4, which follows a cosine annealing schedule for learning rate decay. Hyper-parameter analysis is in Appendix C. |