TL;DR
GoRela introduces a novel motion forecasting framework that achieves viewpoint invariance by reasoning about pairwise relative geometric relationships between agents and map elements. This approach allows for a shared, highly efficient scene encoding that generalizes better to novel scenarios and requires significantly less training data than traditional methods.
What Problem Does This Paper Address?
Predicting where other cars and pedestrians will move is a "make or break" task for self-driving vehicles (SDVs). Currently, researchers face a frustrating trade-off:
- Accuracy vs. Speed: The most accurate models process the world from the perspective of each individual agent. If there are 50 agents, you run the model 50 times—which is too slow for real-time driving.
- Efficiency vs. Generalization: To save time, some models use a single "shared" viewpoint (like the SDV's current position). However, these models often struggle to generalize. If the SDV performs an unusual maneuver, like a sharp U-turn, the "view" changes in a way the model hasn't seen before, leading to poor predictions.
The authors of GoRela tackle this "viewpoint dilemma" by creating a system that is mathematically invariant to the coordinate frame, meaning it doesn't matter where the "camera" is—the relative geometry remains the same.
Methods
The core innovation is PairPose (Pair-wise relative positional encoding) and a Heterogeneous GNN (Graph Neural Network).
- PairPose: Instead of using absolute (x, y) coordinates, GoRela represents the world using the distances and angles between entities (e.g., the distance from Car A to Lane B and the difference in their headings). By using sine and cosine functions of these relative values, the input becomes "viewpoint-invariant."
- Heterogeneous Message Passing (HMP): The scene is treated as a graph where nodes can be agents, lane segments, or crosswalks. The model passes "messages" between these nodes. Crucially, the weights of these messages change depending on the type of relationship (e.g., car-to-car vs. car-to-lane).
- Shared Encoding & Caching: Because the encoding is relative, the heavy lifting of processing the map can be done offline and cached. On the road, the SDV only needs to update the agent-related parts of the graph, drastically reducing latency.
- Goal-Based Decoder: Instead of just drawing a line, the model predicts a "goal" (an endpoint) on the lane graph and then fills in the trajectory. A "greedy goal sampler" ensures the model provides a diverse set of possible futures (e.g., one path for going straight, one for turning).
Datasets
The researchers evaluated GoRela on two major benchmarks:
- Argoverse 2: A massive, industry-standard dataset for urban motion forecasting containing 250,000 scenarios.
- HighwaySim: A new, high-speed highway dataset introduced by the authors to test the model's performance in scenarios where vehicles move much faster and over larger areas than in city driving.
Key Insights and Ideas
- The Power of "Relative": By focusing on how objects relate to each other rather than where they are in a global grid, the model effectively "shrinks" the problem space. It doesn't have to learn what a "left turn" looks like at every possible angle; it only learns the geometry of the turn itself.
- Extreme Sample Efficiency: GoRela is remarkably "data-frugal." The authors found it could outperform baseline models while using 90% less training data. This is a massive advantage given how expensive it is to label high-quality driving data.
- Constant Runtime: Unlike per-agent models that slow down as the scene gets crowded, GoRela’s shared encoding means its runtime stays nearly constant regardless of how many actors are on the road.
- Map-Awareness: By anchoring predictions to the lane graph, the model is much less likely to predict "impossible" trajectories, such as a car driving sideways across a median.
Key Results
- SOTA Performance: On the Argoverse 2 leaderboard, GoRela outperformed existing published methods like THOMAS, particularly in the BrierMinFDE metric (which measures both the accuracy of the path and the confidence of the prediction).
- Highway Superiority: In the HighwaySim tests, GoRela showed a massive improvement in Cross Track Error (CTE), indicating it understands lane boundaries much better than models using global coordinate frames.
- Robustness: When the researchers artificially rotated the scene to "confuse" the models, standard baselines saw their error rates spike. GoRela’s error remained perfectly flat, proving its viewpoint invariance.
Limitations and Future Work
While GoRela is highly effective, the authors noted some specific failure modes in their qualitative analysis:
- Complex Maneuvers: The model occasionally predicts an agent will change lanes when they actually intend to stay in their lane.
- Geometric Precision: In some left-turn scenarios, the model "cut corners" too aggressively compared to the ground truth.
Future work could involve refining the trajectory completion module to better handle these subtle geometric nuances or integrating more complex social cues between agents.
Summary updated