Parameter Identification for Multirobot Systems Using Optimization Based Controllers (Extended Version)

Jaskaran Singh Grover, Changliu Liu, Katia P. Sycara

TL;DR

This paper tackles the challenge of an adversary inferring a multirobot system's task parameters (like goals or controller gains) by only observing robot positions. They develop a framework to determine when this identification is possible, showing that increased robot-to-robot interactions make it harder by obscuring task-specific movements.

What Problem Does This Paper Address?

In multirobot systems, robots often coordinate to achieve tasks while avoiding collisions. This paper focuses on the security and robustness of such systems by asking: can an external observer, by simply tracking the positions of robots, figure out the underlying parameters of their tasks (e.g., their intended destinations or how aggressively they avoid each other)? This is crucial because if an adversary can infer these parameters, they might be able to disrupt the robots' mission. The core difficulty lies in the fact that the robots use optimization-based controllers, which means their movements are implicitly tied to task parameters, making it hard to derive explicit relationships needed for traditional parameter identification methods.

Methods

The authors leverage the concept of Persistency of Excitation (PE) from system identification, which essentially means the system's inputs (or in this case, observed states) need to be "rich enough" over time for accurate parameter estimation. To overcome the implicit dependence of robot control on task parameters, they employ:

  • Duality Theory: This allows them to transform the problem of optimizing control inputs under inequality constraints (collision avoidance) into a more manageable form.
  • Karush-Kuhn-Tucker (KKT) Conditions: These are used to analyze the optimization problem solved by each robot to generate its control.
  • Singular Value Decomposition (SVD): By analyzing the active collision avoidance constraints (those that are actually being enforced at a given moment), they derive an explicit relationship between the robot's control output and its task parameters. This explicit relation is key to applying PE conditions.
  • Observer-Based Parameter Identification: They adapt existing algorithms (like an adaptive observer) that provide conditions for guaranteed convergence of parameter estimates.

The core idea is to classify the robot's dynamics based on the number and linear dependency of active collision constraints, which then dictates whether the task parameters are identifiable.

Datasets

Not discussed in the paper. The evaluation is performed using numerical simulations with varying geometric configurations of robots and obstacles.

Key Insights and Ideas

  • Interaction is Key (for Obfuscation): The more robots an "ego" robot needs to actively avoid, the more its motion is dictated by collision avoidance rather than its primary task. This "hides" the task parameters from an observer.
  • Implicit vs. Explicit Dependence: The main hurdle is the implicit nature of parameter dependence in optimization-based control. By using KKT conditions and SVD on active constraints, they create an explicit link, enabling PE analysis.
  • Constraint Geometry Matters: The number of active constraints and their linear dependency (or independence) directly determines identifiability.
  • Zero or One Active Constraint: Parameters are generally identifiable.
  • Two or More Linearly Independent Active Constraints: Parameters become unidentifiable because the control is fully determined by the constraints, leaving no room for task parameter influence.
  • Two or More Linearly Dependent Active Constraints: Identifiability depends on the specific geometric arrangement and whether the "representative" constraint's orientation is time-varying.

Key Results

The paper derives necessary conditions for parameter identification based on the number and nature of active collision avoidance constraints. These conditions are then validated through simulations using an adaptive observer and an unscented Kalman filter (UKF) for goal estimation.

  • Successful Identification: When the derived conditions predict identifiability (e.g., few active constraints), both the adaptive observer and UKF successfully converge to the true task parameters (goals in the simulations).
  • Failed Identification: When the conditions predict non-identifiability (e.g., multiple independent active constraints), both estimators fail to converge, demonstrating the practical relevance of the theoretical conditions.
  • Qualitative Agreement: The simulation results align with the intuition that increased robot interaction (more active obstacles) makes parameter identification more difficult.

Limitations and Future Work

  • Decoupled Identification: The current analysis focuses on identifying one parameter (goal or gain) at a time, assuming the other is known. Joint identification of multiple parameters is mentioned as a straightforward extension but not fully explored.
  • Specific Controller: While the framework is generalizable, the analysis is demonstrated using a specific Control Barrier Function (CBF)-based Quadratic Program (QP) controller.
  • Adversarial Model: The paper assumes an "adversarial observer" who passively monitors positions. More sophisticated adversaries with active capabilities are not considered.

Future work includes:

  • Extending the framework to joint parameter identification.
  • Applying the theory to infer parameters of other multirobot tasks.
  • Investigating parameter identification for human tasks and how robots can influence this process for human-robot interaction.

Summary updated

Abstract

This paper considers the problem of parameter identification for a multirobot system. We wish to understand when is it feasible for an adversarial observer to reverse-engineer the parameters of tasks being performed by a team of robots by simply observing their positions. We address this question by using the concept of persistency of excitation from system identification. Each robot in the team uses optimization-based controllers for mediating between task satisfaction and collision avoidance. These controllers exhibit an implicit dependence on the task's parameters which poses a hurdle for deriving necessary conditions for parameter identification, since such conditions usually require an explicit relation. We address this bottleneck by using duality theory and SVD of active collision avoidance constraints and derive an explicit relation between each robot's task parameters and its control inputs. This allows us to derive the main necessary conditions for successful identification which agree with our intuition. We demonstrate the importance of these conditions through numerical simulations by using (a) an adaptive observer and (b) an unscented Kalman filter for goal estimation in various geometric settings. These simulations show that under circumstances where parameter inference is supposed to be infeasible per our conditions, both these estimators fail and likewise when it is feasible, both converge to the true parameters. Videos of these results are available at https://bit.ly/3kQYj5J.
BibTeX
@inproceedings{grover2021parameter,
  author = {Jaskaran Singh Grover and Changliu Liu and Katia P. Sycara},
  title = {{Parameter Identification for Multirobot Systems Using Optimization Based Controllers (Extended Version)}},
  booktitle = {MRS},
  year = {2021},
  url = {https://openreview.net/forum?id=4-Dk9pJdNZ},
}
Figure