File size: 1,419 Bytes
cdb11ff | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 | ---
license: apache-2.0
pipeline_tag: other
---
# FlowSim: Rendering-Free Closed-Loop Simulation for Evaluating End-to-End Driving Systems
This model implements the **FlowSim** approach described in [FlowSim: Rendering-Free Closed-Loop Simulation for Evaluating End-to-End Driving Systems](https://huggingface.co/papers/2610.18030).
FlowSim skips image rendering and perception re-encoding in closed-loop evaluation. It predicts the E2E model's internal latent representation for the next frame from the previous latent and the target frame's symbolic scene conditions (agents, map, ego status). The frozen downstream modules (planning, detection, segmentation, etc.) then run on the predicted latent unchanged.
**Key components:**
- Latent transition model: previous latent + target-frame scene conditions → target latent.
- Explicit cross-frame agent association for stable transitions and rollouts.
- Variable-horizon (VH) simulation to avoid accumulating intermediate rollout errors.
Two independent instantiations are provided:
- **`flowsim_navsim`** (sparse) – latent TransFuser (LTF) on the NAVSIM dataset.
- **`flowsim_uniad`** (dense) – UniAD on the nuScenes dataset.
**Project page:** https://hanfeng-wu.github.io/flowsim_page/
**Code:** https://github.com/hanfeng-wu/FlowSim
For pretrained checkpoints and usage, please refer to the [GitHub repository](https://github.com/hanfeng-wu/FlowSim). |