Scenario Diffusion: Controllable Driving Scenario Generation With Diffusion
Authors: Ethan Pronovost, Meghana Reddy Ganesina, Noureldin Hendy, Zeyu Wang, Andres Morales, Kai Wang, Nick Roy
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Scenario Diffusion at generating driving scenarios conditioned on only the map, and with additional conditioning tokens as well. Finally, we provide an analysis of the generalization capabilities of our model across geographical regions, showing that our diffusion-based approach has sufficient expressive capacity to model diverse traffic patterns. 4 Experiments |
| Researcher Affiliation | Industry | Ethan Pronovost Meghana Reddy Ganesina Noureldin Hendy Zeyu Wang Andres Morales Kai Wang Nicholas Roy Zoox {epronovost,mrganesina,nhendy,zewang,andres,kai,nroy}@zoox.com |
| Pseudocode | No | The paper describes the model architecture and processes using text and figures (e.g., Figure 2), but does not provide specific pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing source code for the described methodology or a link to a code repository. |
| Open Datasets | Yes | Datasets We use two datasets in this work. The Argoverse 2 motion forecasting dataset [36] contains 250,000 driving scenarios over 6 geographical regions representing 763 hours of total driving. We use the provided training and validation splits. More information about this dataset can be found in Appendix A.1. |
| Dataset Splits | Yes | We use the provided training and validation splits. |
| Hardware Specification | Yes | We use 2 RTX A6000 GPUs with a batch size of 16 per GPU. We use 4 RTX A6000 GPUs with a batch size of 64 per GPU. |
| Software Dependencies | No | The paper mentions specific algorithms and optimizers used (e.g., Adam optimizer [19], Adam W optimizer [21], EDM diffusion algorithm [18]) but does not list specific version numbers for general software dependencies or libraries (like Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | The scenario autoencoder is trained using the Adam optimizer [19], initial learning rate of 1e-4, and weight decay of 1e-4. ... The KL regularization loss LKL is weighted by 0.1 compared to the reconstruction loss described above. ... We set σmin = 0.02, σmax = 20, σdata = 0.5, ρ = 7, Pµ = 0.5, and Pσ = 1. ... The diffusion model is trained with an Adam W optimizer [21], an initial learning rate of 3e-4, and weight decay of 1e-5. |