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..
VR-Drive: Viewpoint-Robust End-to-End Driving with Feed-Forward 3D Gaussian Splatting
Authors: Hoonhee Cho, Jae-Young Kang, Giwon Lee, Hyemin Yang, Heejun Park, Seokwoo Jung, Kuk-Jin Yoon
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our results demonstrate that VR-Drive is a scalable and robust solution for the real-world deployment of end-to-end autonomous driving systems. Our results demonstrate that VR-Drive is a scalable and robust solution... Table 1 shows the performance of E2E-AD models on both original and novel views... We conducted an ablation study on each module, as shown in Table 2. ... Table 4 shows the closed-loop evaluation results on the Town05-Nov benchmark. |
| Researcher Affiliation | Collaboration | Hoonhee Cho1 Jae-Young Kang1 Giwon Lee1 Hyemin Yang1 Heejun Park1 Seokwoo Jung2 Kuk-Jin Yoon1 1 KAIST 2 42dot |
| Pseudocode | No | The paper describes the framework and methods in narrative text and uses figures to illustrate the architecture (Figure 2, Figure 3). It does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | We do not provide the code and dataset at the time of submission, but we will provide open access to the code and dataset after acceptance. |
| Open Datasets | Yes | We use the nu Scenes [1] benchmark, which is widely used in recent E2E-AD works [38, 26, 5]... We use the CARLA 0.9.10.1 simulator [15] for closed-loop testing. |
| Dataset Splits | Yes | Specifically, we sample 20% of sequences from Town05 Short to construct Town05-Nov, which serves as our novel-view evaluation set. For training data, we follow Transfuser [46] and collect samples using the autopilot, but only from original viewpoints. |
| Hardware Specification | Yes | On an NVIDIA TITAN RTX, each sequence took over 8 hours to train, and the total time of optimization and rendering took more than 3 weeks. |
| Software Dependencies | No | The paper mentions CARLA 0.9.10.1 simulator but does not list specific software dependencies like programming languages or libraries with version numbers. |
| Experiment Setup | Yes | During training, we rendered random novel view images with pitch in the range [ 10 , 5 ], height in [ 0.7m, 1.0m], and depth in [ 0.2m, 1.0m], which broadly covers the test configurations... The loss functions consist of various tasks. For motion prediction and planning, we apply the winner-takes-all strategy [34]. In the planning task, an extra regression loss is introduced to handle ego status. For classification, we utilize focal loss [37], while L1 loss is used for regression in both detection and mapping tasks. Furthermore, L1 loss is also employed for depth estimation. Additionally, we incorporate the viewpoint-consistent distillation loss. We also use a rendering loss for scene reconstruction, as described below. |