Secure Perception-driven Control of Mobile Robots using Chaotic Encryption ACC21 TAC24

Mobile robots integrate heterogeneous devices for embedded sensing, mobile computing and real-time control. These devices exchange information via on-board communication medium. However, the adoption of new technologies brings a wide spectrum of privacy and security issues. This research specifically considers two classes of attacks. One is passive attacks on intra-robot communication, which can be launched to eavesdrop confidential information during data transmission. The other one is active attacks against machine learning systems at test time. Our proposed method is evaluated using a task of perception-driven path tracking in Unreal Engine, where robust path t tracking  and secure image transmission are shown.

Distributed Safe Learning and Planning for Multi-robot Systems  CDC22 ArXiv

Robots operating in the real world are usually accompanied by unknown disturbances. In order to guarantee the safety of the robots as well as mission completion, it is crucial for the robots being able to adapt to the disturbances online and update their motion plans accordingly.  This research considers the problem of online multi-robot motion planning with general nonlinear dynamics subject to unknown external disturbances. We propose dSLAP, the distributed Safe Learning And Planning framework, where the robots collect streaming data to online learn about the disturbances, use the learned model to compute a set of safe actions that avoid collisions against the learning uncertainty, and then choose an action that balances between reaching the goals and actively exploring the disturbances.

Federated Generalizable Reinforcement Learning Automatica24, ACC23, ArXiv

Although classic reinforcement learning methods can deal with complex environments, the agents struggle to generalize their experiences to new environments. This research focuses on the generalization of reinforcement learning, that is, obtaining a control policy which performs well in new environments unseen during training. We propose a novel framework to tackle the challenge of robot motion planning with zero-shot generalization in the presence of distributed data across multiple learning entities. A network of learners aim to collaboratively learn a single control policy which can safely drive a robot to goal regions in different environments without data collection and policy adaptation during policy execution.

Lightweight  Distributed  Gaussian  Process  Regression TAC24 ACC20 ArXiv

However, limited resources challenge implementation of the algorithms on physical agents. In this work, we study the problem where a group of agents aim to collaboratively learn a common static latent function through streaming data. We propose a lightweight distributed Gaussian process regression (GPR) algorithm that is cognizant of agents’ limited capabilities in communication, computation and memory. We show that limited inter-agent communication improves learning performances in the sense of Pareto.

Byzantine-resilient federated Gaussian process regression NeurIPS22

Federated learning provides a promising paradigm for training large volumes of distributed data efficiently without large communication overhead or sharing raw data. However, federated learning faces significant security challenges. For example, Byzantine attacks compromise data owners such that they deviate from expected training processes. In this paper, we study Byzantine-resilient federated online learning for Gaussian process regression (GPR). We develop a Byzantine-resilient federated GPR algorithm that allows a cloud and a group of agents to collaboratively learn a latent function and improve the learning performances where some agents are subject to Byzantine attacks.

Multi-robot-assisted Human Crowd Control.  ACC23

In this work, the interactions between the robots and the humans are modeled by a social force model. A novel optimization problem is formulated to synthesize a controller with closed-form expression.  Specifically, by using Lyapunov stability theory, the convergence of the humans to the exit is imposed as hard constraints, and the convergence of the robots to the exit is imposed as soft constraints.  Sufficient conditions for global asymptotic stability are established for the humans and the robots.

Multi-robot-assisted Human Crowd Evacuation using Navigation Velocity Fields.  TCST24 CDC22 ArXiv

In this paper, we explore the role of two-scale models, particularly the hydrodynamic models, as a tool for designing efficient robot-guided crowd evacuation systems. We start with modeling the microscopic behaviors of humans using social force models and incorporating the robots’ guidance as additional navigation social forces. These individual-based human models include human-human and human-robot interactions and unknown social forces to account for unpredicted environmental change (e.g., dynamic obstacles). Then, we derive the corresponding macroscopic hydrodynamic model which represents the temporal and spatial evolution of the crowd density and flow velocity under the navigation of the robots. The control design for the robots is divided into two parts: position control and direction control.  We prove the stability of the proposed evacuation algorithm and perform a series of simulations involving unknown dynamic obstacles to validate the performance of the algorithm