Research map / Vision-language and multimodal / Multimodal LLMs
EgoDyn-Bench: Evaluating Ego-Motion Understanding in Vision-Centric Foundation Models for Autonomous Driving
Finn Rasmus Schäfer, Yuan Gao, Dingrui Wang, Thomas Stauner, Stephan Günnemann, Mattia Piccinini, Sebastian Schmidt, Johannes Betz
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| provide explicit ego-motion trajectory text (Summary) instead of vision only | Qwen3-VL-8B · EgoDyn-Bench | balanced accuracy: 38.9 -> 59.6 | Improved |
| add visual frames to text-only trajectory input under best encoding (ablation) | Qwen3-VL-8B · EgoDyn-Bench | balanced accuracy: +2.6pp | No clear effect |
| add Timeseries kinematic encoding instead of Summary (ablation) | Qwen3-VL-8B · EgoDyn-Bench trajectory encoding | balanced accuracy: 54.7 -> 62.2 | Improved |
| use Coordinates encoding instead of Summary (ablation) | Qwen3-VL-8B · EgoDyn-Bench trajectory encoding | balanced accuracy: 54.7 -> 46.6 | Hurt |
| use Full (Timeseries + Coordinates) encoding instead of Summary (ablation) | Qwen3-VL-8B · EgoDyn-Bench trajectory encoding | balanced accuracy: 54.7 -> 61.3 | Improved |
| add a single static visual frame to no-visual input (ablation) | Qwen3-VL-8B · EgoDyn-Bench vision-only | WPCR: 20.0 -> 97.4 | Improved |
| add multiple temporally ordered frames beyond a single static frame (ablation) | Qwen3-VL-8B · EgoDyn-Bench vision-only | balanced accuracy: 39.0 (shuffled) vs 38.9 (ordered) | No clear effect |
| increase temporal resolution to 10 FPS (ablation) | Qwen3-VL · EgoDyn-Bench vision-only | balanced accuracy: +0.4pp | No clear effect |
| scale up to closed-source model (gemini-3-pro-preview) from best open 8B model | vision-only VLMs · EgoDyn-Bench vision-only | balanced accuracy: 39.9 -> 47.0 | No clear effect |
| use domain-specific driving VLA instead of general open-source VLM of similar size | RoboTron-Drive · EgoDyn-Bench vision-only | balanced accuracy: 38.6 (RoboTron-Drive) vs 38.9 (Qwen3-VL-8B) | No clear effect |
| use Visual Odometry geometric baseline instead of VLMs | GPT-5.1 · EgoDyn-Bench vision-only (subset) | balanced accuracy: 55.1 -> 63.8 | Improved |
| evaluate on real, simulated or style-transferred domain (ablation) | Qwen3-VL-8B · EgoDyn-Bench domain gap | balanced accuracy: 40.0 / 39.7 / 41.0 | No clear effect |
Closest papers by meaning
- Explicit Geometric Chain-of-Thought for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.89)
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.89)
- QUANTIPHY: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models (CVPR 2026 · similarity 0.89)
- LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.89)
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.89)
- SpaceDrive: Infusing Spatial Awareness into VLM-based Autonomous Driving (CVPR 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
