Research map / Label-efficient and robust learning / Recognition and long-tail
Is Ego Status All You Need for Open-Loop End-to-End Autonomous Driving?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that ego status dominates open-loop nuScenes planning, proposes a Curb Collision Rate metric, and introduces BEV-Planner, a perception-annotation-free baseline that matches SOTA L2/collision scores.
Problem
Open-loop end-to-end autonomous driving on nuScenes may be dominated by ego status and simple straight-driving scenarios, and current L2 and collision metrics fail to capture planning quality, so reported state-of-the-art rankings may be misleading.
Approach
Argues that ego status (velocity, acceleration, yaw, command) drives planning on nuScenes; introduces Curb Collision Rate (CCR) to check road adherence; and proposes BEV-Planner, a simple cross-attention planner over temporally concatenated BEV features trained with only an L1 trajectory loss, with no perception annotations.
Key results
- Ego-MLP, using only ego status, reaches average L2 0.35 m and collision 0.37% on nuScenes, on par with UniAD and VAD.
- Blanking all camera inputs to VAD-Base leaves average L2 at 0.46 m versus 0.37 m with images, while altering ego velocity strongly changes trajectories.
- BEV-Planner++ (with ego status) reaches average L2 0.35 m and collision 0.34%, but BEV-Planner without ego status has worse CCR (4.26%) than UniAD (1.72%) and VAD (2.47%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adding ego status (velocity, acceleration, yaw) as planner input, with no perception | Ego-MLP · nuScenes open-loop planning | avg L2 (m): Ego-MLP 0.35 vs BEV-Planner 0.55 (no ego status) | Helped |
| adding ego status to the BEV-Planner (BEV-Planner++) (ablation) | BEV-Planner · nuScenes open-loop planning | avg L2 (m): 0.55 -> 0.35 | Helped |
| adding map perception task to BEV-Planner (ablation) | BEV-Planner · nuScenes open-loop planning | avg L2 (m) and collision: L2 0.55 -> 0.96 | Hurt |
| adding map perception task to BEV-Planner (ablation) | BEV-Planner · nuScenes open-loop planning | avg CCR (%): CCR 4.26 -> 2.60 | Helped |
| removing ego status / camera input from VAD (blanking images) (ablation) | VAD-Base · nuScenes open-loop planning | avg L2 (m): 0.37 -> 0.46 with blank images | No clear effect |
| setting ego velocity input to zero in VAD (ablation) | VAD-Base · nuScenes open-loop planning | avg L2 (m): 0.37 -> 6.16 | Hurt |
| switching evaluation to a curb collision rate (CCR) metric (ablation) | UniAD post-processing (P.P.) · nuScenes open-loop planning | avg CCR (%): CCR 1.72 without P.P. -> 7.83 with P.P. | Hurt |
| removing UniAD post-processing (ablation) | UniAD · nuScenes open-loop planning | avg collision (%): collision 0.51 with P.P. -> 0.62 without P.P. | Hurt |
Surprising findings
- Blanking all camera input to VAD-Base barely degrades planning L2 and collision, indicating planning is dominated by ego status.
- UniAD's post-processing reduces collision rate but substantially increases curb collisions (CCR 7.83% vs 1.72%).
Limitations
- BEV-Planner performs poorly on the new CCR metric because it uses no map information.
- CCR is imperfect since some annotated nuScenes road boundaries are traversable and collision rate is unreliable when ego leaves the BEV area.
- BEV-Planner is explicitly not intended for real-world deployment due to lack of constraints.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes
Metrics: L2 distance collision rate curb collision rate (CCR) NDS mAP
Baselines: UniAD VAD ST-P3 AD-MLP Ego-MLP GoStright
Builds on: AD-MLP BEVFormer UniAD VAD
Closest papers by meaning
- Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving (NeurIPS 2024 · similarity 0.87)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.86)
- DriveHierarchy: A Benchmark for Diagnosing VLM Driving Capabilities from Open-Loop Understanding to Closed-Loop Execution (NeurIPS 2026 · similarity 0.85)
- Enhancing End-to-End Autonomous Driving with Latent World Model (ICLR 2025 · similarity 0.84)
- NAVSIM: Data-Driven Non-Reactive Autonomous Vehicle Simulation and Benchmarking (NeurIPS 2024 · similarity 0.84)
- Leveraging Driver Field-of-View for Multimodal Ego-Trajectory Prediction (ICLR 2025 · similarity 0.84)
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.