Residual-Signal Protocol (RSP)
Evaluating whether learned interaction representations explain trajectory-forecasting errors beyond physical and scenario-based factors
Status: Submitted to CoRL 2026 (under review).
Overview
The Residual-Signal Protocol (RSP) is an evaluation framework for determining whether learned target–neighbor interaction representations capture information that genuinely helps explain trajectory-forecasting errors beyond simple physical interaction metrics and scenario covariates.
Rather than relying only on aggregate forecasting metrics, RSP examines whether a representation provides meaningful explanatory signal after accounting for established baseline factors.
DyadCode
As part of the framework, I developed DyadCode, a learned discrete representation of pairwise driving interactions. DyadCode captures recurring interaction patterns between a target vehicle and its neighboring agents.
Its explanatory signal is evaluated against random, shuffled, and K-means-based controls using robustness tests, PET/TTC information ablations, and representation-quality analyses.
Evaluation
RSP was evaluated on the Argoverse 2 Motion Forecasting Dataset using forecasting outputs from MTR and QCNet. The evaluation examines:
- whether learned representations explain forecasting errors beyond baseline factors;
- whether the observed signal remains stable under robustness checks;
- how different representations rank difficult interactions; and
- how representation behavior changes across interaction conditions.
Technologies
PyTorch, Argoverse 2, MTR, QCNet, representation evaluation, statistical analysis, and HPC/GPU computing.