Reasoning Multi-Agent Behavioral Topology for Interactive Autonomous Driving
Authors: Haochen Liu, Li Chen, Yu Qiao, Chen Lv, Hongyang Li
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive verification on large-scale real-world datasets, including nu Plan and WOMD, demonstrates that Be Top achieves state-of-the-art performance in both prediction and planning tasks. |
| Researcher Affiliation | Collaboration | Haochen Liu1,2 Li Chen2,3 Yu Qiao2 Chen Lv1 Hongyang Li2,3 1 Nanyang Technological University 2 Shanghai AI Lab 3 University of Hong Kong |
| Pseudocode | No | The paper describes the model architecture and processes using diagrams and textual descriptions, but it does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Code and model is available at https://github.com/Open Drive Lab/Be Top. |
| Open Datasets | Yes | Extensive verification on large-scale real-world datasets, including nu Plan and WOMD, demonstrates that Be Top achieves state-of-the-art performance in both prediction and planning tasks. Data for nu Plan [80] and WOMD [35] are complied with CC-BY-NC 4.0 licence and Apache License 2.0; |
| Dataset Splits | Yes | For planning tasks in nu Plan, there are in total 1M training cases with 8s horizons. 8,300 separated testing set are chosen by Test14-Hard and Test14-Random benchmarks [73] for hard-core and general driving scenes. With further demands verifying maneuvers under interactive cases, we build the Test14-Inter benchmark filtering 1,340 scenes by testing set. The motion prediction tasks in WOMD share 487k training scenarios, with 44k validation and 44k testing set separately partitioned under two challenges: |
| Hardware Specification | Yes | Be Top Net for both prediction and planning tasks are trained in end-to-end manners by Adam W optimizer with 4 NVIDIA A100 GPUs. |
| Software Dependencies | No | Be Top Net for both prediction and planning tasks are trained in end-to-end manners by Adam W optimizer with 4 NVIDIA A100 GPUs. |
| Experiment Setup | Yes | The learning rate is configured as 1e 4 scheduled with the multi-step reduction strategy. The planning model is trained by 25 epochs with a batch size of 128, while the prediction task is trained with 30 epochs with a batch of 256. |