Resilient Active Information Acquisition With Teams of Robots

Brent Schlotfeldt, Vasileios Tzoumas, George J. Pappas

TL;DR

Multi-robot teams often rely on their sensors to map environments or track targets, but what happens if an adversary temporarily blinds some of them? This paper introduces RAIN (Resilient Active Information acquisitioN), a receding-horizon planning algorithm that anticipates worst-case sensor failures, ensuring the robot team continues to gather critical information even when under attack.

What Problem Does This Paper Address?

In robotics, Active Information Acquisition (AIA) is the process of purposefully moving robots around an environment to maximize the data their sensors collect (e.g., mapping an unknown building or tracking moving targets). Standard AIA algorithms assume perfect sensor operation.

However, in adversarial or hazardous environments, sensors can fail or be attacked—for example, a smoke cloud might temporarily block the cameras of several drones. If a team's trajectory plan doesn't account for this, losing a few key robots can completely compromise the mission. While a team could theoretically just "replan" the moment an attack is detected, real-world algorithmic and computational delays make instantaneous replanning impossible. Therefore, robots need to generate trajectory plans that are inherently resilient to a certain number of simultaneous sensor failures during the window of time when replanning is unavailable.

Methods

The authors formulate the problem as a sequential mixed-integer optimization task: maximize the information gathered while minimizing the impact of a worst-case attack that disables up to $\alpha$ robots.

To solve this in real-time, they propose RAIN, a receding horizon (Model Predictive Control-style) algorithm. At the core of RAIN is a subroutine called Robust Trajectory Planning (RTP).

RTP uses a clever two-step "bait" strategy to handle the computationally heavy worst-case attack scenario:

  1. The Bait: RTP identifies a subset of $\alpha$ robots that would individually gather the most information. It assumes the attacker is smart and will target these specific robots, effectively treating them as "bait."
  2. The Backup: RTP then plans the trajectories for the remaining robots assuming the bait robots' sensors are completely disabled.

To solve the underlying multi-robot routing math, RTP relies on coordinate descent (a standard approximation algorithm). The authors also provide rigorous mathematical proofs bounding the suboptimality of RTP using concepts of "curvature" from submodular combinatorial optimization.

Datasets

The paper does not use standard machine learning datasets. Instead, the authors evaluate their algorithms in three simulated environments:

  1. Multi-Target Tracking: A 2D simulation tracking moving targets with unicycle-model robots.
  2. Occupancy Grid Mapping: 2D simulated mapping tasks using a square obstacle map and a corridor map.
  3. Persistent Surveillance: A 3D simulation built on Unity and ROS (Robot Operating System) using the "Camp Lejeune" outdoor building environment provided by ARL DCIST.

Key Insights and Ideas

  • The Power of "Bait": By explicitly sacrificing the expected gains of the highest-performing robots during the planning phase, the algorithm forces the rest of the team to adopt highly robust, overlapping coverage strategies.
  • Submodularity in Robotics: Information gathering often has diminishing returns (e.g., two robots looking at the same spot don't gather twice the info). This property is called submodularity. The authors cleverly use the "curvature" of these submodular functions to mathematically guarantee that their approximation algorithm won't fail arbitrarily badly, even in the worst-case scenario.
  • Worst-Case Planning Generalizes: Planning for a highly intelligent, worst-case attacker also makes the system highly robust to random system failures.

Key Results

RAIN was compared against a state-of-the-art, non-resilient baseline (standard coordinate descent) across three tasks:

  • Multi-Target Tracking: RAIN dramatically reduced the peak Root Mean Square Error (RMSE) of target locations. Depending on the configuration, RAIN achieved peak errors that were 2 to 30 times lower than the baseline, especially when the number of attacks ($\alpha$) was high.
  • Occupancy Grid Mapping: RAIN consistently achieved lower map entropy (meaning higher certainty/better maps) under both worst-case and random attack models.
  • Persistent Surveillance: Even when the team was allowed to replan at every single time step (giving the baseline the best possible chance to react to attacks), RAIN still improved the frequency of landmark observations by ~24%.

Limitations and Future Work

  • Centralized Communication: The current framework assumes all robots can communicate perfectly with each other (or a central planner) at all times to share measurements and coordinate descent steps.
  • Theoretical Tightness: While the authors prove that RTP's performance is bounded, they note that whether these are the tightest possible mathematical bounds remains an open problem in optimization theory.
  • Future Work: The authors suggest extending this framework to distributed settings, where robots must plan resiliently using only local communication with their immediate neighbors.

Summary updated

Abstract

Emerging applications of collaborative autonomy, such as multitarget tracking , unknown map exploration , and persistent surveillance , require robots plan paths to navigate an environment while maximizing the information collected via on-board sensors. In this article, we consider such information acquisition tasks but in adversarial environments, where attacks may temporarily disable the robots’ sensors. We propose the first receding horizon algorithm, aiming for robust and adaptive multirobot planning against any number of attacks, which we call Resilient Active Information acquisitioN ( RAIN ). RAIN calls, in an online fashion, a robust trajectory planning ( RTP ) subroutine that plans attack-robust control inputs over a look-ahead planning horizon. We quantify RTP ’s performance by bounding its suboptimality. We base our theoretical analysis on notions of curvature introduced in combinatorial optimization. We evaluate RAIN in three information acquisition scenarios: multitarget tracking , occupancy grid mapping , and persistent surveillance . The scenarios are simulated in C++ and a unity-based simulator. In all simulations, RAIN runs in real time, and exhibits superior performance against a state-of-the-art baseline information acquisition algorithm, even in the presence of a high number of attacks. We also demonstrate RAIN ’s robustness and effectiveness against varying models of attacks (worst case and random), as well as varying replanning rates.
BibTeX
@article{schlotfeldt2022resilient,
  author = {Brent Schlotfeldt and Vasileios Tzoumas and George J. Pappas},
  title = {{Resilient Active Information Acquisition With Teams of Robots}},
  journal = {IEEE Trans. Robotics},
  year = {2022},
  url = {https://openreview.net/forum?id=c3vb2sfM5X},
}

Loading related content...

Figure