Research map / Vision-language and multimodal / Human and embodied modeling
Int2Planner: An Intention-based Multi-modal Motion Planner for Integrated Prediction and Planning
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.
Proposes Int2Planner, a transformer-based joint prediction-and-planning model that samples multi-modal route intention points from the navigation route to condition ego-vehicle trajectory generation, evaluated on a private dataset, nuPlan, and real-world vehicle tests.
Problem
Motion planning for autonomous driving must handle uncertainty from interactions with surrounding agents, but prediction and planning are usually treated separately, and static or clustered intention points do not reflect the ego vehicle's route-determined intention.
Approach
Replace static/clustered intention points with route intention points resampled at equal distance along the primary and secondary route polylines, generating a multi-modal planning trajectory per intention point within a jointly trained prediction-and-planning transformer decoder.
Key results
- On nuPlan Test14-hard, Int2Planner (learning-only) reaches overall 0.7476 and the hybrid variant reaches 0.7679, with the hybrid claimed to surpass PDM-Hybrid by 7%.
- On the private dataset, route intention points (All) reduce planning ADE from 0.4788 (cluster intention) to 0.3946 and FDE from 1.3436 to 1.0877.
- Removing integrated prediction lowers Test14-hard NR-CL from 0.6971 to 0.6784 (per Table ablation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace clustered intention points with route-sampled intention points (ablation) | Int2Planner, private validation · Private validation | Plan ADE: 0.4788 -> {0.4141} | Helped |
| sample from primary and secondary routes rather than only the primary route (ablation) | Int2Planner, private validation · Private validation | Plan ADE: {0.4141} -> 0.3946 | Helped |
| use route embedding with route intention points (ablation) | Int2Planner · Test14-hard | OL: 0.8606 -> 0.8673 | Helped |
| add integrated prediction (ablation) | Int2Planner · Test14-hard | NR-CL: 0.6784 -> 0.6971 | Helped |
| increase decoder iterations from K=1 to K=6 (ablation) | Int2Planner · Test14-hard | NR-CL: 0.6154 -> 0.6971 | Helped |
| increase decoder iterations from K=6 to K=9 (ablation) | Int2Planner · Test14-hard | OL: 0.8673 -> 0.8736 | Helped |
| increase decoder iterations from K=6 to K=9 (ablation) | Int2Planner · Test14-hard | NR-CL: 0.6971 -> \textbf{}0.6871 | Hurt |
| increase decoder iterations from K=6 to K=9 (ablation) | Int2Planner · Test14-hard | R-CL: 0.6784 -> \textbf{}0.6764 | Hurt |
| use the output of decoder layer k=6 rather than k=1 (ablation) | Int2Planner · Test14-hard | NR-CL: 0.5651 -> 0.6971 | Helped |
| add rule-based post-processing | Int2Planner · Test14-hard | Overall: 0.7476 -> 0.7679 | Helped |
Surprising findings
- Cluster-based intention points give better prediction minADE/minFDE than route intention points on the private dataset (0.5654 vs 0.6010 minADE6), while route intentions dominate planning.
Limitations
- Highest-confidence trajectory is used for closed-loop and real-world tests, which may not be optimal in complex scenarios.
- Learning-only model can produce unsafe behavior without rule-based post-processing; real-world tests required human operators.
Experimental setup
Datasets: private dataset (680,964 scenarios) nuPlan
Benchmarks: nuPlan Val14 nuPlan Test14-hard private dataset
Metrics: Overall score OL score NR-CL score R-CL score ADE FDE minADE6 minFDE6
Baselines: IDM PlanCNN GC-PGP PlanTF PDM-Hybrid GameFormer Planner
Builds on: MTR GameFormer DenseTNT
Closest papers by meaning
- AlignDrive: Aligned Lateral-Longitudinal Planning for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.86)
- Diffusion-Based Planning for Autonomous Driving with Flexible Guidance (ICLR 2025 · similarity 0.86)
- Intend, Reflect, Refine: An Adaptive Multimodal Reflection Framework for Autonomous Driving (NeurIPS 2026 · similarity 0.85)
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.84)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.84)
- Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling (NeurIPS 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.