Research map / Time series / Causal and generative time series
Motion Forecasting in Continuous 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.
Proposes RealMotion, a continuous-driving motion forecasting framework that reorganizes Argoverse scenes into sequential sub-scenes and adds scene-context and agent-trajectory memory streams to relay past information across time.
Problem
Existing motion forecasting methods process each driving scene independently, ignoring temporal relationships between successive scenes encountered during continuous driving, which makes forecasting suboptimal and inefficient in practice.
Approach
Reorganize existing benchmark samples into continuous sub-scene sequences, then propagate information across time through a scene context stream (cross-attention over historical scene features with motion-aware layer norm alignment) and an agent trajectory stream (memory bank of past predictions used to refine current trajectories).
Key results
- On Argoverse 2 test, RealMotion achieves minFDE1 3.93 and minADE1 1.59 versus QCNet's 4.30 and 1.69, and minFDE6 1.24 vs 1.29 (single-agent).
- On Argoverse 1 validation, RealMotion reports minADE6 0.61, minFDE6 0.91, MR6 0.07, best among listed methods on minADE6 and MR6.
- RealMotion (offline) runs at 62ms with 2.9M parameters on RTX 3090 versus QCNet at 94ms with 7.7M parameters; ablation shows the full model improves minFDE6 from 1.423 (no components) to 1.312.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace independent forecasting with RealMotion continuous forecasting | Motion forecasting · Argoverse 2 test set | minFDE_1: 4.42 -> 3.93 | Helped |
| Reorganize scenes into continuous data (ablation) | RealMotion-I · Argoverse 2 validation set | minFDE_1: 4.499 -> 4.397 | Helped |
| Add scene context stream (ablation) | RealMotion with continuous data · Argoverse 2 validation set | minFDE_1: 4.397 -> 4.129 | Helped |
| Add agent trajectory stream (ablation) | RealMotion with continuous data · Argoverse 2 validation set | minFDE_1: 4.397 -> 4.194 | Helped |
| Add both scene context and agent trajectory streams (ablation) | RealMotion with continuous data · Argoverse 2 validation set | minFDE_1: 4.397 -> 4.091 | Helped |
| Add context feature alignment (ablation) | RealMotion · Argoverse 2 validation set | minFDE_6: 1.334 -> 1.328 | Helped |
| Add trajectory feature alignment (ablation) | RealMotion · Argoverse 2 validation set | minFDE_6: 1.334 -> 1.326 | Helped |
| Use both context and trajectory feature alignment (ablation) | RealMotion · Argoverse 2 validation set | minFDE_6: 1.334 -> 1.324 | Helped |
| Add a second alignment module after the first (ablation) | RealMotion · Argoverse 2 validation set | minFDE_6: +0.002 | No clear effect |
| Add trajectory embedding (ablation) | RealMotion with both alignment modules · Argoverse 2 validation set | minFDE_6: 1.324 -> 1.312 | Helped |
| Increase gradient computation steps from 1 to 2 (ablation) | RealMotion · Argoverse 2 validation set | minFDE_6: 1.420 -> 1.341 | Helped |
| Increase gradient computation steps from 2 to 3 (ablation) | RealMotion · Argoverse 2 validation set | minFDE_6: 1.341 -> 1.312 | Helped |
Surprising findings
- The data reorganization alone, without any new streams, already yields a noticeable gain over the independent baseline.
- Splitting trajectories into 5 segments gave limited improvement over 3 segments, indicating longer segment sequences add learning burden.
Limitations
- Requires a sufficient number of historical frames, so it is not applicable to short-history datasets such as Waymo Open Dataset.
- Existing datasets provide limited historical information unlike real-world continuous driving, restricting the sequential design's benefits.
- Failure cases include complex intersection turns and subjective behaviors like parking.
Experimental setup
Datasets: Argoverse 1 Argoverse 2
Benchmarks: Argoverse 2 test (leaderboard) Argoverse 2 validation Argoverse 1 validation Argoverse 2 multi-agent test
Metrics: minADE minFDE MR b-minFDE avgMinFDE avgMinADE actorMR latency
Baselines: QCNet HPTR Forecast-MAE GANet Gnet HDGT THOMAS GoRela SIMPL ProphNet FJMP HiVT
Builds on: QCNet HPTR Forecast-MAE
Closest papers by meaning
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.89)
- Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · similarity 0.88)
- Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.88)
- HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention (CVPR 2024 · similarity 0.87)
- SmartPretrain: Model-Agnostic and Dataset-Agnostic Representation Learning for Motion Prediction (ICLR 2025 · similarity 0.87)
- UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning (NeurIPS 2025 · similarity 0.87)
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.