Research map / Architectures and efficiency / RNNs and normalization
PARA-Drive: Parallelized Architecture for Real-time 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.
Systematically explores the design space of end-to-end modular AV stacks on nuScenes and proposes PARA-Drive, a fully parallel architecture co-training mapping, motion, occupancy and planning heads over a shared BEV feature map.
Problem
End-to-end yet modular autonomous driving stacks differ widely in module choice, placement, and inter-module information flow, and prior evaluation protocols on nuScenes are inconsistent, so the value of these design choices is unclear.
Approach
Build a flexible framework on UniAD to toggle modules and connections, derive insights (BEV features with auxiliary tasks suffice for planning; parallel co-training works), and propose PARA-Drive where planning depends only on the shared BEV features so non-planning modules can be deactivated at runtime for speed.
Key results
- PARA-Drive reduces L2 error and collision rate versus VAD and UniAD on nuScenes val under the standardized protocol (Ave1,2,3s-style averages; e.g., 0.26 vs 0.37 collision and 0.6568 vs 0.9474 L2 Aveall for UniAD).
- PARA-Drive achieves about 2.77x runtime speedup over UniAD-base by deactivating all modules except backbone and planning head.
- With R101 backbone, PARA-Drive's perception and prediction performance is on par or slightly better than UniAD (e.g., IoU-n 63.6 vs 62.3, VPQ-n 55.6 vs 52.8).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PARA-Drive instead of VAD | End-to-end driving on nuScenes val · nuScenes val | L2 error Aveall (m): 0.7830 -> 0.5574 | Helped |
| Use PARA-Drive instead of UniAD | End-to-end driving on nuScenes targeted scenarios · nuScenes targeted scenarios | L2 error Aveall (m): 0.9935 -> 0.9082 | Helped |
| Add ego vehicle states to PARA-Drive (ablation) | PARA-Drive · nuScenes targeted scenarios | L2 error Aveall (m): 0.9082 -> 0.7018 | Helped |
| Remove Map-Mot query and Occ-Plan test-time optimization (ablation) | UniAD · nuScenes val | L2 error Aveall (m): 0.83 -> 0.53 | Helped |
| Add Map-Plan BEV connection (ablation) | Improved baseline · nuScenes val | L2 error Aveall (m): 0.53 -> 0.56 | Hurt |
| Add Map-Plan query connection (ablation) | Improved baseline · nuScenes val | L2 error Aveall (m): 0.53 -> 0.58 | Hurt |
| Add Occ-Plan query connection (ablation) | Improved baseline · nuScenes val | L2 error Aveall (m): 0.53 -> 0.63 | Hurt |
| Remove BEV features from planning (ablation) | Improved baseline · nuScenes val | L2 error Aveall (m): 0.53 -> 4.66 | Hurt |
| Pass motion queries instead of BEV features to planning (ablation) | Improved baseline without BEV features · nuScenes val | L2 error Aveall (m): 4.66 -> 0.54 | Helped |
| Remove online mapping (ablation) | Improved baseline · nuScenes val | offlane rate (%): 1.20 -> 3.03 | Hurt |
| Remove occupancy prediction (ablation) | Improved baseline · nuScenes val | L2 error Aveall (m): 0.5734 -> 0.8174 | Hurt |
| Remove motion prediction (ablation) | Improved baseline · nuScenes val | L2 error Aveall (m): 0.5734 -> 0.8561 | Hurt |
Surprising findings
- Using only the BEV feature map for planning, with proper co-training, can reach state-of-the-art planning performance without explicit upstream dependencies.
- Common evaluation inconsistencies in UniAD, VAD and AD-MLP protocols (time averaging, pedestrian filtering, frame masking, first-frame noise) can reverse reported rankings between VAD and UniAD.
Limitations
- Results limited to open-loop evaluation on nuScenes; closed-loop experiments left for future work.
- Map compliance and targeted-scenario metrics depend on the authors' standardized protocol, which differs from prior evaluations.
- Planning gains from CAN bus and history trajectories become marginal when BEV co-training is used.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes
Metrics: L2 error collision rate map compliance (offroad, offlane) mAP NDS AMOTA minADE IoU-lane IoU-road IoU-n
Baselines: UniAD VAD AD-MLP OccNet
Builds on: UniAD BEVFormer Panoptic SegFormer VAD
Closest papers by meaning
- DriveTransformer: Unified Transformer for Scalable End-to-End Autonomous Driving (ICLR 2025 · similarity 0.86)
- DriveHierarchy: A Benchmark for Diagnosing VLM Driving Capabilities from Open-Loop Understanding to Closed-Loop Execution (NeurIPS 2026 · similarity 0.86)
- Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.85)
- Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving (NeurIPS 2024 · similarity 0.85)
- FIVE-VLA: Fast and EffectIVE Closed-Loop Autonomous Driving with Recurrent Action Memory (NeurIPS 2026 · similarity 0.84)
- End-to-End Autonomous Driving Through V2X Cooperation (AAAI 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.