motion forecasting

**Motion Forecasting** is a **broader generalization of trajectory prediction** — predicting the future state (position, velocity, pose, intention) of dynamic agents in an environment, critical for safety-critical autonomous decision making. **What Is Motion Forecasting?** - **Scope**: Includes Trajectory (where), Pose (body language), and Semantics (lane changes). - **Context**: heavily relies on the static environment (HD Maps, road geometry). - **Uncertainty**: A key requirement is outputting confidence intervals or multiple hypothesis modes. **Why It Matters** - **Collision Avoidance**: The primary safety layer for AV stacks (Waymo, Tesla FSD). - **Interactive Planning**: "If I merge left, will the car behind me slow down?" (Game Theoretic planning). **Techniques** - **VectorNet**: Representing maps and agent paths as vectors. - **LaneGCN**: Using Graph Convolutional Networks to model lane connectivity. - **Interaction Transformers**: Attention over both time (history) and social space (other agents). **Motion Forecasting** is **predictive empathy for robots** — anticipating what others will do so the robot can be a good citizen of the road.