Abstract
Autonomous robots often encounter challenging situations where their control policies fail and an expert human operator must briefly intervene, e.g., through teleoperation. In settings where multiple robots act in separate environments, a single human operator can manage a fleet of robots by identifying and teleoperating one robot at any given time. The key challenge is that users have limited attention: as the number of robots increases, users lose the ability to decide which robot requires teleoperation the most. Our goal is to automate this decision, thereby enabling users to supervise more robots than their attention would normally allow for. Our insight is that we can model the user's choice of which robot to control as an approximately optimal decision that maximizes the user's utility function. We learn a model of the user's preferences from observations of the user's choices in easy settings with a few robots, and use it in challenging settings with more robots to automatically identify which robot the user would most likely choose to control, if they were able to evaluate the states of all robots at all times. We run simulation experiments and a user study with twelve participants that show our method can be used to assist users in performing a simulated navigation task. We also run a hardware demonstration that illustrates how our method can be applied to a real-world mobile robot navigation task.
Introduction
In the paradigm of "sliding autonomy," robots handle routine tasks independently, while human operators intervene via teleoperation for edge cases—such as navigating complex crowds or recovering from sensor failures. While effective for single-robot systems, this approach faces a significant bottleneck when scaled to fleets. As the number of robots grows, the primary constraint shifts from the operator's control skill to their attention.
Managing a large fleet becomes a high-stakes game of "Whack-a-Mole." An operator cannot simultaneously monitor dozens of robot states to determine which one requires immediate assistance. This "attention gap" often leads to wasted effort on robots performing adequately while others remain stuck. This paper proposes a method to automate the selection process, effectively pointing the "human spotlight" where it is needed most.
Methodology: Scaling Human Attention
The proposed framework scales human autonomy through a four-phase process:
- Initial Policy Training: A base autonomous policy ($\pi_R$) is established using standard imitation learning from human demonstrations on a single robot.
- Learning the Scoring Model: The system observes a user managing a small, manageable fleet (e.g., 4 robots). By applying the Luce choice model, the system treats human interventions as relative preferences (e.g., "Robot A needs help more than Robot B"). A neural network is trained to estimate a scoring function ($\hat{\phi}$) that maps robot states to an intervention priority value.
- Assisted Choice: In large-scale deployments (e.g., 12+ robots), the system uses the learned $\hat{\phi}$ to automatically switch the operator's view and control to the robot with the highest predicted need.
- Policy Improvement: Human interventions in these critical states provide high-quality demonstration data, which is used to further refine the autonomous policies.
Experimental Setup
The approach was validated across three distinct environments:
- ViZDoom Simulation: A custom navigation task requiring robots to avoid hazards and collect items.
- Human User Study: Twelve participants managed 12 robots in ViZDoom, comparing manual switching against the proposed assisted switching.
- Hardware Demonstration: Real-world deployment using NVIDIA JetBot mobile robots navigating a laboratory environment with physical obstacles, using OptiTrack for positioning.
Results and Key Insights
The experimental results demonstrate that modeling relative preferences is significantly more effective than absolute binary classifications of "failure" states.
- Predictive Accuracy: In simulations, the learned scoring model achieved 79% agreement with ground-truth intervention preferences, compared to only 32% for a baseline binary classifier.
- Team Performance: The user study revealed that "Assisted Choice" significantly improved team performance ($p < 0.01$). Participants reported feeling more successful and found the task less cognitively demanding.
- Hardware Validation: The system successfully identified high-priority states in real-world settings, specifically near obstacles and boundaries where autonomous policies typically struggle.
A critical advantage of this method is the "Expert Data Flywheel." By directing human attention to the most challenging states, the system naturally curates the most informative data for imitation learning, creating a virtuous cycle of autonomous improvement.
Conclusion and Future Work
This work demonstrates that human intervention preferences are consistent enough to be mapped and scaled from small to large fleets. However, limitations remain: for extremely complex tasks, learning the scoring function may become as difficult as the task itself. Furthermore, the model currently assumes stationary human preferences. Future research could explore active learning to handle uncertainty and adaptive models that account for shifting human strategies over extended operations.
Summary updated