Abstract
The task of motion forecasting is critical for self-driving vehicles (SDVs) to plan safe maneuvers. Modern approaches reason about the map, agents' past trajectories, and their interactions to produce accurate forecasts. However, the predominant per-agent encoding approach is computationally expensive, while shared-frame encoding is often sample inefficient and vulnerable to domain shift. We propose GoRela, an efficient shared encoding for all agents and the map that leverages pair-wise relative positional encodings in a heterogeneous spatial graph. This parameterization is invariant to scene viewpoint and allows for offline map embedding computation. Our viewpoint-agnostic decoder predicts agent goals on the lane graph to enable diverse and context-aware multimodal prediction. We demonstrate the effectiveness of our approach on the Argoverse 2 benchmark and a novel highway dataset.
Introduction
Predicting the future trajectories of surrounding agents is a fundamental challenge for autonomous driving. Current methodologies often face a "viewpoint dilemma." High-accuracy models typically perform inference from the perspective of each target agent, which scales poorly as the number of agents increases. Conversely, models using a shared coordinate frame (e.g., centered on the SDV) are more efficient but struggle with generalization; unusual SDV maneuvers can create "unseen" viewpoints that degrade performance.
GoRela addresses this by introducing a framework that is mathematically invariant to the coordinate frame. By focusing on relative geometric relationships, the model maintains consistent performance regardless of the scene's global orientation or the SDV's specific pose.
Methodology
The GoRela architecture consists of three main components: relative scene encoding, goal prediction, and trajectory generation.
Pair-wise Relative Positional Encoding (PairPose)
Instead of absolute coordinates, GoRela utilizes PairPose to represent the world through distances and angles between entities. For any two nodes (agents or map elements), the model encodes their relative distance, relative heading, and the angle of the vector between them. Using trigonometric functions of these relative values ensures the input is viewpoint-invariant.
Heterogeneous Graph Neural Network
The scene is represented as a heterogeneous graph where nodes represent agents, lane segments, or crosswalks. We employ Heterogeneous Message Passing (HMP) layers to process these relationships. This allows the model to learn distinct interaction patterns, such as how a car responds to another car versus how it follows a lane boundary.
Efficiency and Goal-Based Decoding
Because the map-to-map relationships are static and relative, map embeddings can be precomputed offline and cached. During real-time inference, the system only updates agent-related edges, significantly reducing latency. The decoder then predicts a "goal heatmap" on the lane graph, from which goals are sampled to generate diverse, map-constrained trajectories.
Results
GoRela was evaluated on the Argoverse 2 urban dataset and a new HighwaySim dataset introduced by the authors.
- Performance and Robustness: GoRela achieved state-of-the-art results on Argoverse 2, particularly in the BrierMinFDE metric. Crucially, while baseline models' error rates spiked when the input scene was rotated, GoRela's performance remained perfectly stable, confirming its viewpoint invariance.
- Sample Efficiency: A key finding is GoRela's extreme sample efficiency. The model outperformed standard baselines while using only 10% of the training data, demonstrating that relative geometric reasoning simplifies the learning task.
- Highway Navigation: In high-speed scenarios (HighwaySim), GoRela showed significant improvements in Cross Track Error (CTE), indicating a superior understanding of lane-following constraints compared to global-frame models.
Conclusion
GoRela demonstrates that viewpoint invariance is a powerful inductive bias for motion forecasting. By shifting from global coordinates to relative geometric relationships, we achieve a model that is both computationally efficient and highly robust to novel scenarios. Future work will focus on refining trajectory precision in complex maneuvers, such as tight left turns where current models may occasionally cut corners.
Summary updated