Research map / Label-efficient and robust learning / Knowledge distillation
LaKD: Length-agnostic Knowledge Distillation for Trajectory Prediction with Any Length Observations
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 LaKD, a plug-and-play length-agnostic knowledge distillation framework with dynamic soft-masking that lets trajectory predictors (HiVT, QCNet) handle observed trajectories of any length.
Problem
Trajectory prediction models degrade sharply when the observed history length is short or variable, yet real driving often requires immediate predictions from few observed points.
Approach
A single shared encoder distills latent features from the currently better-predicted observation length to worse ones (direction chosen dynamically by prediction error), while a gradient-based dynamic soft-masking of neuron units protects important units from knowledge collision.
Key results
- On Argoverse 1 with QCNet backbone (K=6), LaKD reduces minADE from 0.5791 (Orig) and 0.5411 (FLN) to 0.5240.
- On Argoverse 2 with QCNet backbone (K=1), LaKD achieves minADE 1.6574 vs 1.6940 for FLN and 2.1006 for Orig.
- Ablation on Argoverse 1 with HiVT (K=1): removing the dynamic soft-masking and distillation raises minADE from 1.3317 to 1.4189 (random masking only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply LaKD to QCNet | QCNet · Argoverse 1 | minADE (K=6): 0.5411 -> 0.5240 | Helped |
| Apply LaKD to QCNet | QCNet · nuScenes | minADE (K=6): 1.4653 -> 1.4594 | Helped |
| Apply LaKD to QCNet | QCNet · Argoverse 2 | minADE (K=6): 0.7370 -> 0.7258 | Helped |
| Add random masking (ablation) | HiVT on Argoverse 1 · Argoverse 1 | minADE (K=1): 1.4733 -> 1.4189 | Helped |
| Add length-agnostic knowledge distillation (ablation) | HiVT with random masking on Argoverse 1 · Argoverse 1 | minADE (K=1): 1.4189 -> 1.3619 | Helped |
| Add dynamic soft-masking (ablation) | HiVT with random masking and LaKD on Argoverse 1 · Argoverse 1 | minADE (K=1): 1.3619 -> 1.3317 | Helped |
| Vary mask number M from 2 to 6 (ablation) | LaKD with HiVT · Argoverse 1 | minADE (K=1): 1.3457; 1.3317; 1.3414; 1.3563; 1.3486 | No clear effect |
Surprising findings
- Longer observed trajectories do not always yield better predictions; extra history can introduce interference that hurts accuracy.
Limitations
- How to determine 'good' versus 'bad' trajectories is heuristic, based on distance to ground truth.
- Single run results with no error bars due to training cost.
- Code release deferred until acceptance.
Experimental setup
Datasets: Argoverse 1 nuScenes Argoverse 2
Benchmarks: Argoverse 1 nuScenes Argoverse 2
Metrics: minADE minFDE MR
Baselines: HiVT-Orig HiVT-RM HiVT-DTO HiVT-FLN QCNet-Orig QCNet-RM QCNet-DTO QCNet-FLN
Builds on: HiVT QCNet DTO FlexiLength Network
Closest papers by meaning
- NEST: A Neuromodulated Small-world Hypergraph Trajectory Prediction Model for Autonomous Driving (AAAI 2025 · similarity 0.88)
- KnowLCP: Knowledge Augmented Lane Change Prediction for Autonomous Driving (AAAI 2026 · similarity 0.88)
- On the Road to Portability: Compressing End-to-End Motion Planner for Autonomous Driving (CVPR 2024 · similarity 0.87)
- HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention (CVPR 2024 · similarity 0.86)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.86)
- T4P: Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory (CVPR 2024 · similarity 0.85)
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.