The BATmobile: Towards a Bayesian Automated Taxi

Jeff Forbes, Timothy Huang, Keiji Kanazawa, Stuart Russell

Introduction

The BATmobile project represents a significant shift in autonomous driving research, moving away from the "platooning" concepts of the mid-90s—which relied on restricted, instrumented lanes—and into the unpredictable environment of normal highway traffic. By leveraging Dynamic Probabilistic Networks (DPNs), the system maintains a "belief state" that accounts for noisy sensors, occluded vehicles, and the hidden intentions of other drivers to make safe, real-time decisions.

The Challenge of Normal Traffic

In the 1990s, projects like PATH focused on groups of cars traveling in segregated lanes under centralized control. While efficient, this approach requires massive infrastructure investment and fails to address the complexities of the "real world." The authors of the BAT (Bayesian Automated Taxi) project tackle the harder problem of driving in normal traffic, characterized by:

  • Sensor Noise: Imperfect GPS and vision data mean positions are always approximate.
  • Partial Observability: Other vehicles may be hidden behind trucks, and critical driver states (like distraction or aggression) cannot be observed directly.
  • Stochastic Actions: A car's own maneuvers, such as braking on a slick road, do not always have a guaranteed outcome.

The BAT Architecture

The core of the BAT is a decision-theoretic architecture. Unlike rigid "if-then" rule-based systems, it integrates sensor data over time to estimate hidden variables and calculate the utility of various actions.

The high-level architecture of the BAT system, integrating vision, decision-making, and physical simulation.

Methods: Dynamic Probabilistic Networks

The system relies on several key methodological innovations to handle uncertainty efficiently:

  1. Dynamic Probabilistic Networks (DPNs): These Bayesian networks extend over time, allowing the car to integrate sensor data across multiple "time slices." This is crucial for estimating hidden states, such as a neighboring driver's intent to change lanes.
  2. Temporally Invariant Networks: To maintain real-time performance, the authors "roll up" the network, pre-compiling its structure so it remains consistent across time steps, avoiding expensive re-computations.
  3. State Evolution and Sensor Models: The DPN structure explicitly models how the world changes (State Evolution) and how sensors report that world (Sensor Model).
The DPN structure models state evolution and sensor readings over discrete time steps.
  1. Stochastic Simulation: Using a "survival of the fittest" sampling method—similar to a genetic algorithm—the system focuses its computational resources on the most likely real-world states, even when sensors are highly accurate but the world model is "weak."
  2. Sensor Validation: A dedicated sub-network monitors sensor health, adjusting confidence levels if readings deviate wildly from predictions (e.g., during bad weather).

Key Insights

  • Belief State over Raw Data: Driving requires a probability distribution over the state of the world that integrates history, rather than reacting to isolated sensor readings.
  • Indexical Variables: To manage complexity, the authors use "indexical" nodes (e.g., "Front Clear"). A preprocessing step maps the relevant nearby car to these nodes, keeping the DPN small and efficient.
  • Utility-Based Safety: Recognizing that no policy is 100% safe, the system quantitatively weighs risks against benefits (utility) to choose the best action.

Results and Simulation

The system was tested using SmartPATH, a 2-D physical simulator that generates 3-D rendered video output. It successfully navigated five complex scenarios:

  1. Passing: Deciding to change lanes to maintain target speed.
  2. Defensive Driving: Reacting when the probability of another car cutting in exceeds a threshold.
  3. Obstacle Avoidance: Distinguishing between traffic jams and stalled vehicles.
  4. Aborting Maneuvers: Canceling a lane change if the target space becomes occupied simultaneously by another vehicle.
  5. Merging: Finding gaps in traffic as a lane ends.

Limitations and Future Work

While the BATmobile demonstrated superior handling of uncertainty compared to rule-based systems, several challenges remain. Early implementations faced significant computational overhead, and the high-level decision trees were still largely hand-constructed. Additionally, converting continuous variables like velocity into discrete states often led to a "combinatorial explosion." Moving forward, the authors emphasized the need for real-world video validation to transition beyond the SmartPATH simulator.

Summary updated

BibTeX
@inproceedings{forbes1995batmobile,
  author = {Jeff Forbes and Timothy Huang and Keiji Kanazawa and Stuart Russell},
  title = {{The BATmobile: Towards a Bayesian Automated Taxi}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1995},
}
Figure