CARPAL: Confidence-Aware Intent Recognition for Parallel Autonomy

Xin Huang, Stephen G. McGill, Jonathan A. DeCastro, Luke Fletcher, John J. Leonard, Brian C. Williams, Guy Rosman

TL;DR

CARPAL is a novel multi-task neural network that improves driver assistance systems by predicting not just where a car will go, but how "useful" (safe and intent-aligned) those paths are. By estimating task-specific utility and its uncertainty in real-time, the system can make smarter decisions about when to intervene or issue warnings, significantly reducing false alarms while maintaining high safety standards.

What Problem Does This Paper Address?

Advanced Driver Assistance Systems (ADAS) and "parallel autonomy" (where a computer monitors a human driver) face a major hurdle: trustworthy intervention.

Most current systems use "task-agnostic" confidence measures. For example, a predictor might say, "I'm 70% sure the driver will turn left." However, this doesn't tell the system if that 30% uncertainty actually matters for safety. If the driver is heading toward a parked car, a task-agnostic system might fail to intervene because its prediction uncertainty is high, even if the risk is obvious. Conversely, systems that are too sensitive cause "false alarms," annoying drivers and leading them to disable the safety features.

The challenge is creating a system that understands the downstream consequences of its predictions—specifically regarding safety and the driver’s intended goal—and acts only when it is confident that an intervention is necessary and beneficial.

Methods

The authors propose CARPAL (Confidence-Aware intent Recognition for ParALlel autonomy), which consists of two main components integrated into a single pipeline:

  1. Variational Trajectory Predictor: This uses a CNN (VGG-based) to process camera images and a state encoder for vehicle data (steering, speed, etc.). It outputs a probabilistic distribution of potential future trajectories.
  2. Utility Regressor: This is the core innovation. Instead of just predicting paths, this sub-network regresses the mean ($\mu$) and variance ($\sigma^2$) of the "utility" for both the predicted human path and a potential backup autonomous path.
  3. Utility ($u$) is defined by two factors: Safety (distance to obstacles) and Intent Similarity (how well the path matches what the human likely wants to do).
  4. By regressing these statistics directly, the system avoids the "prohibitively expensive" computational cost of running a full planner and safety evaluator thousands of times per second during live operation.

The system uses a Binary Decision Function to decide whether to intervene. It compares the expected utility of the human driver against a backup planner (a hybrid A* search). If the human's utility is low and the system is confident in its assessment (low variance), it takes control. If the utility is low but uncertainty is high, it issues a warning instead.

Datasets

  • Naturalistic Driving Dataset: The model was trained on 0.6 million samples (roughly 30 hours) of real-world urban driving data collected by a test vehicle in various weather and lighting conditions.
  • Augmented Risky Test Set: Since real-world data is (thankfully) mostly safe, the researchers created "counterfactual" risky scenarios. They did this by:
  • Scaling up the speed of recorded trajectories so they would hit obstacles.
  • Adding "ghost" obstacles (like pedestrians or walls) into the camera feed to simulate sudden hazards.

Key Insights and Ideas

  • Task-Specific Uncertainty: The "aha moment" is that prediction error (how far the predicted path is from the actual path) is less important than utility error. CARPAL cares about whether the safety of the path is certain, not just the exact coordinates of the car.
  • Real-Time Efficiency: Calculating utility and planning backup paths online is slow (~500ms). By training a neural network to approximate these values, CARPAL can generate these statistics in just ~9.25ms, making it viable for high-speed robotics.
  • The "Warning" Logic: By reasoning about the variance ($\sigma^2$) of the utility, the system can distinguish between "I know this is a crash" (Intervene) and "This looks risky, but I'm not totally sure" (Warn). This nuanced approach is key to reducing driver frustration.

Key Results

  • Superior ROC Performance: CARPAL achieved a higher Area Under the Curve (AUC) compared to baselines like Velocity-Based Predictors (VBP) and Accuracy-Based Predictors (ABP). It demonstrated a better balance between Recall (catching risks) and Fall-out (avoiding false alarms).
  • Utility Improvement: In risky scenarios, the system improved the overall utility (safety/intent alignment) by an average of 35.6%.
  • Acausal Parity: The authors showed that their "causal" model (which only sees the past) performs nearly as well as an "acausal" model (which has access to the actual future ground truth), proving the effectiveness of their utility regressor.

Limitations and Future Work

  • Perception Dependency: The system relies on a semantic segmentation stack to identify obstacles. If the perception system fails (e.g., missing a small object), the utility calculation will be wrong.
  • Closed-Loop Testing: The current study was performed on recorded "open-loop" data. The authors note that future work should involve testing in a closed-loop simulation or live vehicle where the system's interventions actually change the car's future state.
  • Ternary Decisions: While the logic allows for warnings, the quantitative analysis focused on binary (intervene vs. not). Future work could formalize a more complex ternary decision-making framework.

Summary updated

Abstract

Predicting driver intentions is a difficult and crucial task for advanced driver assistance systems. Traditional confidence measures on predictions often ignore the way predicted trajectories affect downstream decisions for safe driving. In this letter, we propose a novel multi-task intent recognition neural network that predicts not only probabilistic driver trajectories, but also utility statistics associated with the predictions for a given downstream task. We establish a decision criterion for parallel autonomy that takes into account the role of driver trajectory prediction in real-time decision making by reasoning about estimated task-specific utility statistics. We further improve the robustness of our system by considering uncertainties in downstream planning tasks that may lead to unsafe decisions. We test our online system on a realistic urban driving dataset, and demonstrate its advantage in terms of recall and fall-out metrics compared to baseline methods, and demonstrate its effectiveness in intervention and warning use cases.
BibTeX
@article{huang2021carpal,
  author = {Xin Huang and Stephen G. McGill and Jonathan A. DeCastro and Luke Fletcher and John J. Leonard and Brian C. Williams and Guy Rosman},
  title = {{CARPAL: Confidence-Aware Intent Recognition for Parallel Autonomy}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2021},
  url = {https://openreview.net/forum?id=hOuabdyrGHz},
}
Figure