RESEARCH

Sponsored Projects

A robotic rehabilitation gym can be defined as multiple patients training with multiple robots or passive sensorized devices in a group setting. Recent work with such gyms has shown positive rehabilitation outcomes; furthermore, such gyms allow a single therapist to supervise more than one patient, increasing cost-effectiveness. To allow more effective multipatient supervision in future robotic rehabilitation gyms, this project investigates automated systems that could dynamically assign patients to different robots within a session in order to optimize group rehabilitation outcome. 

Selected publications:

[1] Miller, B. A., Adhikari, B., Jiang, C., & Novak, V. D. (2022). Automated patient-robot assignment for a robotic rehabilitation gym: a simplified simulation model. Journal of NeuroEngineering and Rehabilitation, 19(1), 126. 

[2] Adhikari, B., Bharadwaj, V. R., Miller, B. A., Novak, V. D., & Jiang, C. (2023). Learning Skill Training Schedules From Domain Experts for a Multi-Patient Multi-Robot Rehabilitation Gym. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 31, pp. 4256-4265.  

[3] Miller, B. A., Adhikari, B., Jiang, C., & Novak, V. D. (2023). Automated patient-robot task assignment in a simulated stochastic rehabilitation gym. In 2023 IEEE International Conference on Rehabilitation Robotics (ICORR), pp. 1-6. 

[4] Adhikari, B., Ranashing, S., Miller, B. A., Novak, V. D., & Jiang, C. (2022). Learning Dynamic Patient-Robot Task Assignment and Scheduling for A Robotic Rehabilitation Gym. In 2022 IEEE International Conference on Rehabilitation Robotics (ICORR), pp. 1-6. 

Rehabilitation gym scheduling using mixed-integer nonlinear programming (MINLP) [1]: Two examples of total skill gain over time using four schedule types: disjunctive, time-indexed system, best robot, and switch halfway. Example a) is for group 2 with 6 patients, 5 robots, and 7 time steps. Example b) is for group 1 with 5 patients, 7 robots, and 12 time steps. Our study shows that intelligently moving patients between different rehabilitation robots can improve overall skill acquisition in a multi-patient multi-robot environment.

Learning skill training schedules from domain expert [2]: Violin plots of mean skill gain for 20 patient groups with different scheduling methods in the three rehabilitation gym scenarios. Overall, our method can learn different scheduling behaviors from human experts with accuracies of 75-85%, and these learned behaviors result in similar skill gain to actual expert behaviors.

Multi-Robot Coordinated Planning and Control

1. Multi-robot trajectory optimization via model predictive control: 

Multi-robot motion planning and control has been investigated for decades and is still an active research area due to the growing demand for both performance optimality and safety assurance. In this research, we explored optimization-based methods for coordinated control of multiple robots with optimized control performance and guaranteed safety constraints. We developed two distributed model predictive control (DMPC) methods: 1) a gradient-based algorithm that leverages the alternating direction method of multipliers (ADMM) to decompose the team-level trajectory optimization into subproblems solved by individual robots. The algorithm also incorporates a discrete-time control barrier functions (CBF) as safety constraints to provide formal guarantee of collision avoidance; and 2) a sampling-based method that formulates multi-robot optimal control as probabilistic inference over graphical models and leverages belief propagation to achieve inference via distributed computation. We developed a distributed sampling-based model predictive control (MPC) algorithm based on the proposed framework, which obtains optimal controls via variational inference (VI).

Selected publication:

[1] Jiang, C. (2024). Distributed Sampling-Based Model Predictive Control via Belief Propagation for Multi-Robot Formation Navigation. IEEE Robotics and Automation Letters, vol. 9, no. 4, pp. 3467-3474. DOI: 10.1109/LRA.2024.3368794.

[2] Jiang, C., & Guo, Y. (2023). Incorporating Control Barrier Functions in Distributed Model Predictive Control for Multi-Robot Coordinated Control. IEEE Transactions on Control of Network Systems. DOI: 10.1109/TCNS.2023.3290430. 

A multi-robot team represented as a Markov random field.

An overview of the distributed VI-MPC algorithm [1]

video.mp4

Multi-robot navigation in formation using the distributed VI-MPC algorithm [1]

Incorporating control barrier functions (CBF) method in DMPC for multi-robot coordinated control [2]

2. Learning-enabled multi-robot formation: 

Formation control of multi-robot systems has been extensively studied by model-based methods, where analytic controls are constructed based on the kinematics and/or dynamics model and the communication graphs of multi-robot system. Motivated by remarkable advances of robotic learning techniques, emerging methods for learning-enabled formation control have been developed to advance adaptive and intelligent control of multi-robot systems. This project develops policy learning methods for decentralized formation control using deep neural networks which map a robot’s local observations to control actions. Neural network control policies are trained with imitation learning and our proposed guided policy learning method that enables sample-efficient learning. To address the scalability of learned control policy, an aggregation graph neural network (GNN) is used to model distributed and permutation-invariant inter-robot communication.  

Selected publications:

[1] Jiang, C., Huang, X., & Guo, Y. (2023). End-to-End Decentralized Formation Control Using Graph Neural Network Based Learning Method. Frontiers in Robotics and AI, 10, 1285412. 

[2] Jiang, C., & Guo, Y. (2020). Multi-robot guided policy search for learning decentralized swarm control. IEEE Control Systems Letters, 5(3), 743-748. 

[3] Jiang, C., Chen, Z., & Guo, Y. (2019). Learning decentralized control policies for multi-robot formation. In 2019 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM) (pp. 758-765). 


Snapshots of triangulation formation using a GNN-based policy (9 robots) [1]

Triangulation formation using GNN-based policies for different robot team sizes [1]

Imitation learning of decentralized control policies for multi-robot formation [3]

Experimental results in CoppeliaSim simulator [3]

Robot-Assisted Pedestrian Regulation: Learning optimal human-robot interaction (HRI)

This project investigates an emerging application of assistive robots in pedestrian regulation. We propose a novel robot-assisted pedestrian regulation framework that utilizes the dynamic pedestrian-robot interaction during their collective movements. The insights of the effect of pedestrian-robot interaction on the pedestrian movements and the optimal robot motion for desired pedestrian regulation objectives are provided.  Adaptive dynamic programming (ADP) algorithm and deep reinforcement learning algorithms are designed to learn optimal control of robot motion. The proposed adaptive learning framework is applied to a merging flow scenarios to reduce the risk of crowd disasters. Furthermore, to address the challenge of feature representation of complex human motion dynamics under the effect of HRI, an end-to-end robot motion planner based on deep neural network is proposed and trained using a deep reinforcement learning algorithm. The new approach avoids hand-crafted feature detection and extraction and thus improves the learning capability for complex dynamic problems. 

Selected publications:

[1] Jiang, C., Ni, Z., Guo, Y., & He, H. (2019). Pedestrian flow optimization to reduce the risk of crowd disasters through human–robot interaction. IEEE Transactions on Emerging Topics in Computational Intelligence, 4(3), 298-311. 

[2] Wan, Z., Jiang, C., Fahad, M., Ni, Z., Guo, Y., & He, H. (2018). Robot-assisted pedestrian regulation based on deep reinforcement learning. IEEE Transactions on Cybernetics, 50(4), 1669-1682. 

[3] Jiang, C., Ni, Z., Guo, Y., & He, H. (2017). Learning human–robot interaction for robot-assisted pedestrian flow optimization. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 49(4), 797-813. 

ADP-Based Robot Motion Control for Pedestrian Regulation

Optimization of Merging Pedestrian Flows through Human-Robot Interaction

Pedestrian Regulation Using Deep Reinforcement Learning

Indoor Human Localization: A cooperative localization scheme using robot-smartphone collaboration

Smartphone-based human indoor localization was previously implemented using wireless sensor networks at the cost of sensing infrastructure deployment. Motivated by the increasing research attention on location-aware human-robot interaction, this project studies a robot-assisted human indoor localization scheme using acoustic ranging between a self-localized mobile robot and smartphones from human users. Data from the low-cost Kinect vision sensor are fused with smartphone-based acoustic ranging, and an extended Kalman filter based localization algorithm is developed for real-time dynamic position estimation and tracking. Real robot-smartphone experiments are performed, and performances are evaluated in various indoor environments under different environmental noises and with different human walking speed. Compared with existing indoor smartphone localization methods, the proposed system does not rely on wireless sensing infrastructure, and has comparable localization accuracy with increased flexibility and scalability due to the mobility of the robot. 

Selected publications:

[1] Jiang, C., Fahad, M., Guo, Y., & Chen, Y. (2018). Robot-assisted smartphone localization for human indoor tracking. Robotics and Autonomous Systems, 106, 82-94. 

[2] Jiang, C., Fahad, M., Guo, Y., Yang, J., & Chen, Y. (2014). Robot-assisted human indoor localization using the Kinect sensor and smartphones. In 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (pp. 4083-4089). 

Cooperative localization scheme