Reproduce “Recover to Predict: Progressive Retrospective Learning for Variable-Length Trajectory Prediction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hao Zhou, Lu Qi, Jason Li, Jie Zhang, Yi Liu, Xu Yang, Mingyu Fan, Fei Luo
PRF reconstructs missing trajectory history in successive steps before a shared predictor forecasts vehicle motion, instead of using a separate predictor for each history length. On Argoverse 2 validation with DeMo and 10 observed timesteps, it reports mADE₆/mFDE₆ of 0.617/1.183, versus 0.639/1.262 for DeMo-LaKD. These are lowest-of-six average and endpoint displacement errors; validation does not establish real-world safety.
Key insights
- Progressive alignment of features, decomposing the complex task of bridging large information gaps into smaller, manageable retrospective steps, is more effective than direct one-shot mapping for variable-length trajectory prediction.
- Distilling features by modeling the discrepancy caused by omitted timesteps as a learnable residual, conditioned on scene context, effectively preserves reliable information and imputes omissions.
- Implicit supervision from a historical trajectory recovery module during training (RPM) enhances the learning of robust and accurate feature representations without incurring additional inference costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.