RetroMotion: Retrocausal Motion Forecasting Models are Instructable
RetroMotion, an innovative motion forecasting model, breaks down multi-agent trajectory predictions into both marginal and joint distributions. By utilizing a transformer, it re-encodes marginal distributions and captures pairwise interactions while integrating retrocausal information flow. The approach utilizes compressed exponential power distributions to address positional uncertainty. RetroMotion demonstrated impressive performance in the Waymo Interaction Prediction Challenge and shows adaptability to the Argoverse 2 and V2X-Seq datasets.
Key facts
- Motion forecasts vary in complexity based on agents, scene constraints, and interactions.
- Joint trajectory distributions grow exponentially with agent count.
- RetroMotion decomposes forecasts into marginal and joint distributions.
- A transformer model generates joint distributions via re-encoding marginal distributions and pairwise modeling.
- Retrocausal flow moves information from later marginal points to earlier joint points.
- Positional uncertainty is modeled using compressed exponential power distributions.
- Strong results in Waymo Interaction Prediction Challenge.
- Generalizes to Argoverse 2 and V2X-Seq datasets.
Entities
Institutions
- Waymo
- Argoverse
- V2X-Seq