CMU's Causality-Based CBF Makes Autonomous Driving Less Conservative
New paper infers inter-vehicle causality to avoid overly timid robot behavior.
Safe control in autonomous driving often relies on Control Barrier Functions (CBFs) to prevent collisions. Existing CBF methods either assume fully cooperative vehicles or use worst-case analysis, leading to overly conservative behavior that hampers efficiency. A new paper from Carnegie Mellon University tackles this by embedding causality inference into a Parametric-CBF framework. The key insight: instead of treating all surrounding vehicles as adversarial, the system learns the underlying controllers of neighboring vehicles from observed data, accounting for the mutual influence between them.
This causality-based Parametric-CBF allows an autonomous vehicle to form accurate expectations of other vehicles' next moves and adapt in real time. The result is a safety-critical controller that is both safe and less conservative, especially in dense, interaction-heavy traffic. The paper, accepted at the 2026 IEEE International Conference on Robotics and Automation (ICRA), demonstrates significant improvements in task efficiency without sacrificing safety. The work represents a step toward more natural and fluid autonomous driving in complex multi-agent scenarios.
- Extends Parametric-CBF with explicit causality inference to model inter-vehicle influence
- Avoids worst-case conservative behavior by learning neighboring vehicles' controllers from data
- Accepted at ICRA 2026; improves task efficiency in interaction-intensive scenarios
Why It Matters
Smarter, less timid autonomous vehicles that navigate dense traffic safely without unnecessary overcaution.