Research map / Label-efficient and robust learning / Recognition and long-tail
Adapting to Length Shift: FlexiLength Network for Trajectory Prediction
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.
Identifies the Observation Length Shift in Transformer-based trajectory predictors and proposes FlexiLength Network (FLN), which trains sub-networks on multiple observation lengths with shared weights, KL-based temporal distillation, and per-length positional encoding and layer norm.
Problem
Trajectory prediction models degrade significantly when evaluated at observation lengths different from the length used in training, and retraining per length (Isolated Training) is costly.
Approach
Train one set of shared weights on trajectories truncated to several observation lengths, distill the longest-observation predictive distribution into shorter-length sub-networks via KL loss, and use independent positional encodings and specialized layer norms per length so that a single model serves multiple observation lengths.
Key results
- On nuScenes with AgentFormer, FLN reaches ADE5/FDE5 of 1.92/3.91 at 2 timesteps versus 2.02/4.23 for isolated training, and 1.83/3.78 at 4 timesteps versus 1.86/3.89.
- On Argoverse 1 validation with HiVT-64, FLN achieves ADE6/FDE6 of 0.81/1.25 at 10 timesteps versus 0.92/1.43 for isolated training, with 680K parameters versus 662K per model for isolated training.
- Ablations on nuScenes show removing temporal distillation makes FLN worse than isolated training at 2 and 3 timesteps (2.23/4.39 at 2 timesteps).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply FLN to AgentFormer instead of standard training | AgentFormer · nuScenes, 2–4 timesteps | ADE5: 2.26 -> 1.92 (2 timesteps); 2.00 -> 1.88 (3 timesteps); 1.86 -> 1.83 (4 timesteps) | Helped |
| Apply FLN to HiVT-64 instead of isolated training | HiVT-64 · Argoverse 1 validation, 10/20/30 timesteps | ADE6: 0.92 -> 0.81; 0.81 -> 0.72; 0.69 -> 0.65 | Helped |
| Apply FLN to HiVT-64 instead of isolated training | HiVT-64 · Argoverse 1 validation, 10/20/30 timesteps | FDE6: 1.43 -> 1.25; 1.17 -> 1.08; 1.04 -> 0.98 | Helped |
| Use four training observation lengths (10, 16, 20, 30) rather than three (10, 20, 30) (ablation) | HiVT-64-FLN · Argoverse 1 validation, 10/20/30 timesteps | ADE6: 0.81 -> 0.79; 0.72 -> 0.72; 0.65 -> 0.68 | Mixed |
| Add sub-network weight sharing (ablation) | AFormer-FLN · nuScenes, 2/3/4 timesteps | ADE5: 1.94 -> 1.92; 1.89 -> 1.88; 1.92 -> 1.83 | Mixed |
| Add temporal distillation (KL loss) (ablation) | AFormer-FLN · nuScenes, 2/3/4 timesteps | ADE5: 2.23 -> 1.92; 1.98 -> 1.88; 1.85 -> 1.83 | Mixed |
| Add independent positional encoding (ablation) | AFormer-FLN · nuScenes, 2/3/4 timesteps | ADE5: 2.13 -> 1.92; 2.02 -> 1.88; 1.95 -> 1.83 | Helped |
| Add specialized layer normalization (ablation) | AFormer-FLN · nuScenes, 2/3/4 timesteps | ADE5: 2.12 -> 1.92; 1.98 -> 1.88; 1.93 -> 1.83 | Helped |
| Train with observation lengths 10 and 30 only rather than 10, 20, and 30 (ablation) | HiVT-64-FLN · Argoverse 1 validation, 20 timesteps | ADE6: 0.72 -> 0.84 | Hurt |
Surprising findings
- Simply training with the standard length and evaluating at shorter lengths (AgentFormer) causes a clear performance drop, which motivates the paper.
- Mixed sampling improves shorter-length performance but degrades the standard 4-timestep performance.
Limitations
- FLN increases training time because it processes several input sequences per training iteration.
- Observation lengths longer than the standard length cannot be generated easily due to pre-segmentation of the data.
Experimental setup
Datasets: ETH/UCY nuScenes Argoverse 1
Benchmarks: ETH/UCY nuScenes Argoverse 1
Metrics: ADE FDE
Baselines: AgentFormer HiVT-64 Mixed Sampling Fine-tuning Joint training Isolated Training
Builds on: AgentFormer HiVT Transformer knowledge distillation
Closest papers by meaning
- Adapting to Observation Length of Trajectory Prediction via Contrastive Learning (CVPR 2025 · similarity 0.88)
- LaKD: Length-agnostic Knowledge Distillation for Trajectory Prediction with Any Length Observations (NeurIPS 2024 · similarity 0.84)
- Curvature Tuning: Provable Training-free Model Steering From a Single Parameter (NeurIPS 2025 · similarity 0.82)
- HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention (CVPR 2024 · similarity 0.82)
- GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.81)
- Towards Generalizable Trajectory Prediction using Dual-Level Representation Learning and Adaptive Prompting (CVPR 2025 · similarity 0.81)
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.