Spline-Transformer Motion Predictor

This repository contains the weights for a Transformer-based motion prediction model trained on the Argoverse 2 dataset.

Model Architecture

• Base Architecture: Transformer Encoder with Pre-Layer Normalization
• Output Representation: 6 Bezier Spline Control Points
• Trajectory Generation: Differentiable Bezier Spline Decoder outputs 30 future timesteps (3 seconds at 10Hz)
• Input Representation: 20 past timesteps (2 seconds) of relative (x, y) coordinates
• Embedding Dimension: 768
• Attention Heads: 8
• Encoder Layers: 5

Training Details

• Scale Factor: 50.0 (Inputs and targets are divided by 50.0 before entering the model, and predictions are multiplied by 50.0 for real-world coordinate mapping).
• Loss Function: Smooth Trajectory Loss (combining Huber Loss for ADE/FDE, a Continuity Anchor, and a Kinematic Smoothing Penalty).

Usage

To use these weights, initialize the TransformerMotionPredictor with the parameters listed above and load the state_dict.