CausalAgents: A Robustness Benchmark for Motion Forecasting using Causal Relationships
Proposes CausalAgents benchmark, using removal of non-causal agents to evaluate motion prediction robustness, with 25-38% minADE change observed.
Key Findings
Methodology
This study labels causal agents—those influencing human driving behavior—in WOMD via human annotations. Using these labels, scenes are perturbed by removing non-causal agents, then multiple state-of-the-art models (SceneTransformer, MultiPath++, Wayformer) are evaluated under these perturbations. Two metrics are introduced: per-example absolute minADE change and IoU-based trajectory similarity, quantifying model sensitivity. Experiments show that increasing training data and targeted augmentation by dropping non-causal agents improve robustness.
Key Results
- All models exhibit significant performance drops under non-causal perturbations, with relative minADE changes of 25-38%. For example, SceneTransformer’s mean minADE increases from 0.250m to 0.067m after removing non-causal agents. Sensitivity varies with architecture and scene complexity, with some scenarios showing over 1m increase. Data augmentation reduces average sensitivity by about 30%.
- Model sensitivity correlates with architecture and scene complexity. SceneTransformer marginal models are more robust, while MultiPath++-All shows smaller relative changes. Sensitivity increases with higher vehicle speed and larger fractions of removed non-causal agents. Distance from the ego vehicle also influences robustness, with closer agents having higher impact.
- IoU-based trajectory difference metrics reveal that most predicted trajectories change minimally, but some examples show over 1m deviation, indicating vulnerabilities in certain scenarios. This highlights the importance of causal understanding for safety-critical applications.
Significance
This work pioneers the use of human-annotated causal relationships as a systematic benchmark for evaluating motion prediction robustness. It addresses the critical need for models that can reliably handle scene variations and potential spurious correlations, thereby enhancing safety in autonomous driving. The approach bridges causal reasoning with deep learning evaluation, offering a new pathway to develop more resilient models and safety guarantees in complex traffic environments.
Technical Contribution
The paper introduces a causal-based scene perturbation framework, leveraging human labels to identify and remove non-causal agents. It proposes two novel metrics—per-example minADE change and trajectory IoU—to quantify sensitivity. Extensive evaluation across multiple models demonstrates that training data augmentation and targeted dropping of non-causal agents significantly improve robustness. This integration of causal understanding into robustness assessment offers a new paradigm for safety-critical AI systems.
Novelty
This is the first systematic evaluation of motion prediction models based on human-annotated causal relationships, using agent removal as a perturbation strategy. The dual metrics provide comprehensive sensitivity analysis, surpassing prior approaches that relied solely on statistical noise or synthetic perturbations. The work emphasizes the importance of causal reasoning for safety, setting a new standard for robustness benchmarks in autonomous driving.
Limitations
- Relies on manual causal annotations, which are costly and subjective, limiting scalability. Future work should explore automated causal inference methods.
- Perturbations focus mainly on agent removal, not on more complex scene modifications like dynamic interactions or environmental changes, which could affect generalizability.
- Model robustness improvements are promising but may not fully address extreme scenarios or rare edge cases. Further integration with multi-modal data and multi-task learning is needed.
- Experiments are primarily conducted on static scenes and specific architectures; broader validation across diverse environments and models remains necessary.
Future Work
Future directions include automating causal relationship detection to scale annotations, integrating multi-modal sensor data for richer scene understanding, and applying reinforcement learning combined with causal inference to enhance robustness. Extending the benchmark to dynamic and environmental perturbations will further improve safety guarantees, ultimately advancing autonomous vehicle deployment in complex real-world scenarios.
AI Executive Summary
Autonomous driving relies heavily on accurate motion prediction to ensure safety, but current models often struggle with scene variations and unseen scenarios. Traditional evaluation methods, which typically test models on static datasets, fail to capture the full spectrum of real-world complexities. Recognizing this gap, the authors introduce CausalAgents, a novel benchmark that leverages human-annotated causal relationships within traffic scenes. By identifying agents that influence driving behavior, the study constructs scene perturbations through the systematic removal of non-causal agents, thus simulating potential real-world distractions or irrelevant objects.
The core innovation lies in quantifying how sensitive models are to these perturbations. The authors evaluate multiple state-of-the-art models, including SceneTransformer, MultiPath++, and Wayformer, observing that performance, measured by minADE, can deteriorate by 25-38% under non-causal agent removal. To better understand this sensitivity, they develop two metrics: a per-example absolute change in minADE and an IoU-based trajectory similarity measure. These metrics reveal that models often exhibit significant output shifts, with some examples experiencing deviations exceeding 1 meter.
Importantly, the study demonstrates that increasing training data and applying targeted data augmentation—specifically, randomly dropping non-causal agents—can substantially improve robustness. This suggests that models can learn to ignore irrelevant scene elements, reducing brittleness. The findings highlight the critical importance of causal reasoning in safety-critical AI applications, especially in autonomous vehicles, where mispredictions can lead to accidents.
By releasing the causal agent labels and robustness benchmarks publicly, the authors provide valuable tools for the community. This work paves the way for future research integrating causal inference, multi-modal data, and reinforcement learning to develop more reliable, safe autonomous driving systems capable of handling complex, unpredictable environments.
Deep Dive
Abstract
As machine learning models become increasingly prevalent in motion forecasting for autonomous vehicles (AVs), it is critical to ensure that model predictions are safe and reliable. However, exhaustively collecting and labeling the data necessary to fully test the long tail of rare and challenging scenarios is difficult and expensive. In this work, we construct a new benchmark for evaluating and improving model robustness by applying perturbations to existing data. Specifically, we conduct an extensive labeling effort to identify causal agents, or agents whose presence influences human drivers' behavior in any format, in the Waymo Open Motion Dataset (WOMD), and we use these labels to perturb the data by deleting non-causal agents from the scene. We evaluate a diverse set of state-of-the-art deep-learning model architectures on our proposed benchmark and find that all models exhibit large shifts under even non-causal perturbation: we observe a 25-38% relative change in minADE as compared to the original. We also investigate techniques to improve model robustness, including increasing the training dataset size and using targeted data augmentations that randomly drop non-causal agents throughout training. Finally, we release the causal agent labels (at https://github.com/google-research/causal-agents) as an additional attribute to WOMD and the robustness benchmarks to aid the community in building more reliable and safe deep-learning models for motion forecasting.