Human-Drone Partnerships for Emergency Response

Overview

Small Unmanned Aerial Systems (sUAS) used in emergency response missions, including river search-and-rescue and flood reconnaissance, must operate in dynamic and hazardous environments. To accelerate search tasks, multi-UAV platforms often use Human-on-the-Loop (HoTL) autonomy, allowing drones to navigate and search independently while human operators supervise.

Relying entirely on onboard computer vision models in safety-critical missions introduces specific risks:

1. Deep neural networks can fail when encountering out-of-distribution conditions such as water glare, rapids, fog, low daylight, or precipitation.
2. Vision models often output high confidence scores even when making incorrect predictions under domain shift.
3. A false negative can result in a missed victim swept downstream, while a false positive diverts emergency personnel and rescue equipment.

This research was funded by the National Science Foundation (NSF) under grant CNS:1931962 as part of the DroneResponse project. We developed an adaptive autonomy framework that evaluates the real-time reliability of onboard vision models using uncertainty quantification and covariate shift analysis, temporarily lowering autonomy to engage human operators when reliability drops.

Vision failures in autonomous systems

In a Human-in-the-Loop (HitL) system, human operators inspect every video stream and approve individual actions. When swarms of multiple sUAS operate concurrently, manual oversight creates operator fatigue and processing bottlenecks. In fully autonomous Human-on-the-Loop (HoTL) systems, vision errors can propagate into incorrect actions.

A notable example of computer vision failure occurred in the 2018 Uber ATG vehicle accident, where the perception system detected a pedestrian six seconds before impact but cycled through inconsistent classifications (unknown object, vehicle, bicycle) due to dark clothing, lack of lighting, and unreflective angles, failing to alert the operator in time.

In aerial search-and-rescue over water bodies, vision models encounter related failure modes:

• Environmental covariate shifts: Models trained primarily in clear weather lose recall when exposed to river foam, snow, rain, or low-angle sun glare.
• Sensor noise and motion: Propeller wash, mount vibrations, flight blur, and water spray degrade input frames.
• Visual occlusions: Variations in clothing, water immersion, and river debris obstruct the person being tracked.

To prevent incorrect autonomous actions, aerial systems require mechanisms to detect when their vision models are operating outside their reliable domain.

Adaptive autonomy framework

The system evaluates two metrics at runtime:

1. Confidence Score ($CS$): The raw output probability generated by the vision model for a detection.
2. Loss of Reliability ($LoR$): A binary assessment combining model uncertainty and environmental covariate shift.

+-----------------------------------------------------------------------------------+
|                            DroneResponse Adaptive Autonomy                        |
+-----------------------------------------------------------------------------------+
|                                                                                   |
|  [ Aerial Video Stream ]                                                          |
|           |                                                                       |
|           v                                                                       |
|  +--------------------------------+       +------------------------------------+  |
|  |     Computer Vision Model      |       |       Reliability Evaluator        |  |
|  | (YOLO / DeepLab Object Detect) |       |  (Bayesian Uncertainty + Covariate)|  |
|  +--------------------------------+       +------------------------------------+  |
|           |                                                 |                     |
|           | Detection: CS (Confidence)                      | LoR: TRUE / FALSE   |
|           +-----------------------+-------------------------+                     |
|                                   |                                               |
|                                   v                                               |
|               +---------------------------------------+                           |
|               |      Autonomy Adaptation Engine       |                           |
|               |            (EARS Rules)               |                           |
|               +---------------------------------------+                           |
|                   /                               \                               |
|                  /                                 \                              |
|   [ CS >= High & LoR = FALSE ]        [ CS >= High & LoR = TRUE ]                 |
|                 v                                   v                             |
|    High Autonomy Mode (HoTL)            Temporarily Lower Autonomy (HitL)         |
|  - Autonomous victim tracking          - Raise High-Priority Urgent Alert         |
|  - Auto-adjust flight path             - Stream imagery to Human Supervisor       |
|  - Log action to dashboard             - Prompt: "Confirm Victim" / "Get Views"   |
|                                                                                   |
+-----------------------------------------------------------------------------------+

When $LoR = \text{FALSE}$, the drone operates in High Autonomy Mode (HoTL), autonomously tracking the candidate victim, adjusting camera gimbals, and coordinating search paths. If the operating environment differs substantially from training data or sensor noise spikes ($LoR = \text{TRUE}$), the drone temporarily transitions to Human-in-the-Loop (HitL), raising an alert for operator confirmation.

Reliability estimation pipeline

Model reliability is computed through two components: Bayesian uncertainty estimation and covariate shift analysis.

1. Bayesian uncertainty quantification

Uncertainty stems from two primary sources:

• Data (aleatoric) uncertainty: Sensor noise, camera glare, and physical vibrations, modeled as a Gaussian distribution based on sensor specifications.
• Model (epistemic) uncertainty: Parameter uncertainty from sparse training coverage, estimated via Monte Carlo test-time dropout across network parameters.

Full system uncertainty is calculated by propagating data uncertainty through the network layers. To avoid rapid switching between autonomy levels from brief noise spikes, we use a hysteresis band that quantizes uncertainty into three states:

• Confident: Autonomous tracking proceeds (HoTL).
• Uncertain: Autonomy continues with increased telemetry monitoring.
• No Confidence: Autonomy is reduced, and operator verification is requested ($LoR = \text{TRUE}$).

2. Covariate shift analysis

A vision model can output high confidence on out-of-distribution inputs. We used a generative density model to estimate model performance across continuous environmental covariates:

1. Automated attribute annotation:
   • Weather classifier: An ensemble of binary Support Vector Machines (SVMs) classifying daylight, rain, fog, and snow.
   • Semantic segmentation: A DeepLab network segmenting water bodies, rapids, riverbanks, and vegetation.
2. Pairwise similarity estimation: Constructs an $N^2$ similarity matrix comparing the query image vector $X_k$ against training dataset distributions.

3. Density regression: Calculates the conditional True Positive Rate ($\text{TPR}$) for a given covariate precondition vector $X$ at an acceptable false-positive rate ($\text{fpr}$):

   \[\text{TPR}(\text{fpr} \mid X) = F_M\left(F_{\bar{M}}^{-1}(\text{fpr} \mid X) \;\middle|\; X\right)\]

   Where $F_M$ and $F_{\bar{M}}$ are cumulative distribution functions for matched and non-matched covariate distributions. When $X$ indicates insufficient training coverage, $LoR$ is set to $\text{TRUE}$.

Autonomy adaptation requirements

System requirements were specified using the Easy Approach to Requirements Syntax (EARS):

• loss_of_reliability = TRUE when:

  \[\text{Uncertainty} > \tau_{\text{uncertainty}} \quad \lor \quad \text{Covariate Coverage} < \tau_{\text{coverage}}\]

Adaptation rules:

• AR1 (Confident Autonomy - HoTL):
  When the vision model identifies a candidate victim with $CS \ge \text{detect_threshold}$ and $\text{loss_of_reliability} = \text{FALSE}$, the sUAS autonomously transitions to tracking mode and logs a notification.
• AR2 (Autonomy Reduction - HitL):
  When the vision model identifies a candidate victim with $CS \ge \text{detect_threshold}$ and $\text{loss_of_reliability} = \text{TRUE}$, the sUAS temporarily lowers autonomy, raises an alert on the operator console, and requests confirmation before tracking.
• AR3 (Ambiguous Detection):
  When the vision model detects a potential target with $\text{alert_threshold} \le CS < \text{detect_threshold}$, the sUAS records a spatial marker on the map and continues its search sweep.

Under AR2, the operator can confirm the victim, reject a false alarm, or instruct the drone to obtain more imagery from an alternate altitude or angle.

DroneResponse emergency response platform

The framework was evaluated within DroneResponse, an autonomous multi-UAV system developed at the University of Notre Dame:

• Applications: River search-and-rescue, swift-water monitoring, and flotation device delivery.
• Hardware and streaming: Edge computing modules (such as NVIDIA Jetson) running lightweight object detectors, paired with RTSP video streaming.
• Supervisory interface: A centralized ground control station allowing one operator to monitor multiple autonomous drones.
• Alert prioritization: Dynamic alert budgeting to avoid overwhelming the operator when multiple drones detect candidate targets simultaneously.

Research challenges

The paper outlined five challenges for deploying vision models safely in autonomous robotics:

1. Covariate discovery: Identifying which latent environmental variables impact model reliability without requiring exhaustive manual labeling.
2. Automated attribute extraction: Developing reliable attribute classifiers for lighting, weather, and terrain meta-characteristics.
3. Real-time context monitoring: Combining onboard sensor data, external weather feeds, and visual scene analysis at runtime.
4. Safety assurance cases: Formally verifying that autonomy switching thresholds balance safety requirements against operator workload.
5. Human-robot interaction risks: Mitigating operator complacency from over-trusting autonomous decisions, while avoiding loss of situational awareness during long flights.

Project details and publications

• Role: Graduate Research Assistant and Computer Vision Researcher
• Affiliation: Computer Vision Research Laboratory (CVRL) and Software Engineering Research Laboratory, University of Notre Dame
• Sponsorship: Sponsored by the National Science Foundation (NSF) under grant CNS:1931962.

Publication

• Title: Adaptive Autonomy in Human-on-the-Loop Vision-Based Robotics Systems
• Authors: Sophia Abraham, Zachariah Carmichael, Sreya Banerjee, Rosaura G. VidalMata, Ankit Agrawal, Md Nafee Al Islam, Walter Scheirer, Jane Cleland-Huang
• Venue: 2021 IEEE/ACM 1st Workshop on AI Engineering (WAIN), co-located with ICSE 2021.

• Links: arXiv:2103.15053

  Conference Proceedings