Ego-Agent Trajectory Prediction in Argoverse 2

Project Overview

This project implements a deep learning model for forecasting the future motion of an ego vehicle in complex multi-agent driving scenarios. Built with PyTorch and trained on a modified version of the Argoverse 2 Motion Forecasting Dataset, the model was developed for a Kaggle-hosted course challenge.

Our model is inspired by the ADAPT architecture, and designed to handle long-horizon trajectory prediction by encoding temporal and social context. Specifically, we:

• Use bidirectional LSTM to encode past trajectories for all agents
• Apply multi-head self-attention to model agent-agent interactions
• Predict a coarse future endpoint, refine it, and condition trajectory generation on it
• Detach gradients between stages for training stability

We trained our model on agent-centric motion features: position (x, y), velocity (vx, vy), heading angle, and object type. No map-based information (such as HD lanes, semantic raster, or traffic light states) was provided, highlighting the model’s ability to perform under minimal context assumptions.

• 🔗 View GitHub Repo
• 📄 View Report (PDF)

Model Architecture

Results

The model was evaluated using trajectory-level Mean Squared Error (MSE) between predicted and ground-truth ego-agent trajectories.