Hierarchical Approach to Transfer of Control in Semi-Autonomous Systems

Kyle Hollins Wray, Luis Enrique Pineda, Shlomo ZilbersteinIJCAI 2016

Introduction

As we move toward semi-autonomous vehicles, the "hand-off" problem becomes a critical safety bottleneck. Most current systems treat the human as a backup or use hard-coded logic to switch control. However, humans are unpredictable—they might be distracted, drowsy, or slow to respond to alerts. Existing planning algorithms often fail to explicitly account for the uncertainty and time-sensitivity of transferring control during the execution of a long-term plan. If a car needs to hand control back to a driver before a complex intersection, it needs to know not just that it must hand over control, but how to do so successfully given the driver's current state.

Methodology

The researchers propose a hierarchical approach to bridge the gap between second-to-second interaction and minute-to-minute navigation:

  1. The Micro-Level (TOC POMDP): The transfer of control is modeled as a Partially Observable Markov Decision Process (POMDP). This handles the uncertainty of the human's state (e.g., attention level) based on noisy sensor data. The agent chooses "messages" (visual or auditory alerts) to nudge the human back into the loop, balancing the effectiveness of the alert against the "annoyance" or cost to the user.
  2. The Macro-Level (SSP MDP): The overall route planning is modeled as a Stochastic Shortest Path (SSP) problem. Crucially, it treats the low-level TOC process as an "option" (a macro-action). This allows the high-level planner to understand the probability of a successful hand-off at any given intersection without getting bogged down in the millisecond-level details of the sensors.
Figure 1: Illustration of a planned route involving a Transfer of Control (TOC) from a human-driven segment to an autonomy-capable road.

Safety Guarantees

The authors formally define "Live States"—conditions where at least one actor is capable of acting competently. They provide a theoretical analysis showing that their hierarchical model is a "Strong SAS," meaning it provably avoids "dead-end" states where neither the human nor the AI can safely control the vehicle. A critical component of the TOC model is the ability to "abort" and safely pull over if a human fails to take control.

Experimental Evaluation

The model was evaluated using real-world road network data from OpenStreetMap (OSM) across 10 different cities (including Austin, Boston, NYC, and San Francisco). "Main roads" were designated as autonomy-capable and "back roads" as human-only to simulate current technological limitations. The researchers compared three configurations: a human-only driver (λ), an autonomous-only vehicle (ν), and their collaborative "SAVE" (Semi-Autonomous VEhicle) model (λ & ν).

Key Results

The collaborative model demonstrated significant improvements in both reachability and efficiency across diverse urban environments.

Figure 2: Performance comparison across 10 cities showing reachability (G), autonomy percentage (%), and travel time (T) for human (λ), autonomous (ν), and combined (λ & ν) models.
  • Reachability: The autonomous-only vehicle failed to reach the goal in 7 out of 10 cities because it couldn't navigate "human-only" segments. The collaborative model reached the goal 100% of the time.
  • Autonomy Maximization: In cities like Boston, the collaborative system allowed the vehicle to drive autonomously for 95% of the trip duration, only handing over to the human for the final "last mile" or complex segments.
  • Efficiency: Travel times for the collaborative model were nearly identical to the human-only driver, proving that safety-conscious hand-offs do not necessitate significant delays.

Conclusion

By decoupling the process of transfer from route planning, the authors developed a computationally feasible approach for city-scale navigation. Instead of reacting to failures, the system proactively plans routes based on where successful transfers are most likely. Future work will involve testing with real human drivers in high-fidelity simulators and exploring dynamic updates to road capabilities.

Summary updated

BibTeX
@inproceedings{wray2016hierarchical,
  author = {Kyle Hollins Wray and Luis Enrique Pineda and Shlomo Zilberstein},
  title = {{Hierarchical Approach to Transfer of Control in Semi-Autonomous Systems}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure