Viterbi Faculty Directory

Research Summary

---

I am interested in developing human-centered automation solutions to increase human productivity, reduce human health risks, and enable innovation. Realizing human-centered automation requires robotics and automation technologies to reduce the physical human effort and enable humans to focus their effort on high-value tasks. It also requires augmenting human decision-making capabilities to reduce the probability of making mistakes, enabling creativity, and increasing decision-making speed.

During the early part of my career, I focused on developing computationally efficient geometric reasoning techniques, and integrating these techniques with heuristic search methods to solve manufacturability analysis and process planning problems. As the next logical step, I focused on integrating physics-based simulations into planning algorithms to generate plans that can be safely executed in the real-world. I explored both model simplification and meta-modeling approaches to ensure that planning problems can be solved in real-time and can produce good quality plans. I also examined how uncertainty might influence the risk associated with the generated plans and developed methods to generate risk-informed plans. I also investigated how planning problems that involve both discrete and continuous variables can be solved by combining optimization problems with discrete state space search.

Our recent work is focused on solving sequential decision-making problems to endow robots with capabilities to serve as smart assistants for humans. Many emerging robotics applications require multiple collaborating robots to operate under human supervision. To be useful in such applications, smart robotic assistants will need to (1) program themselves, (2) efficiently learn from the observed performance, (3) safely operate in the presence of uncertainty, (4) appropriately call for help during the execution of challenging tasks, and (5) effectively communicate with humans. We are leveraging a physics-informed machine learning approach to speed up sequential decision-making in robotics applications. We use both active learning methods to enable robots to do efficient exploration to build the word model, and we also use imitation learning and inverse reinforcement learning methods to learn from human demonstrations. We have been training neural networks using a combination of data generated from physical experiments and simulations to enable real-time decision making. We have also developed new learning architectures to ensure that the required physics models are enforced during the training. We are developing methods for robots to operate safely in the presence of uncertainty by generating contingency-aware plans. We are developing computational methods for endowing robots with introspective capabilities so that they can seek help from humans on challenging tasks. We are also exploring the use of augmented reality-based interfaces for enabling robots to elicit human guidance.

To augment human decision-making capabilities, we are interested in developing decision support systems that enable humans to make informed decisions in a timely manner for challenging applications. We are making advances in physics-informed artificial intelligence to enable the realization of decision support systems for applications that involve model uncertainty, complex physics, and fast decision-making speeds. We are also developing a novel framework to combine model-based approaches and data-driven approaches in a consistent and unified manner. This enables us to exploit prior knowledge and augment missing components of models through safe and efficient learning. We are also developing a digital twin based alert system that proactively notifies the human supervisor of possible adverse events and serves as a decision support system for humans to make informed decisions.

Based on previous and on-going research, my group has made the following ten notable contributions:

1. Our patented technology in the area of setup planning for sheet metal bending operations is the basis of Japanese machine tool manufacturer Amada's automated process planning software for robotic press brakes. Software based on this technology is widely used in the sheet metal industry.

2. Our work in-mold assembly processes enables realizing geometrically complex heterogeneous structures in a cost-effective manner. These advances have enabled the use of novel, bio-inspired design concepts in applications such as robotics, bio-medical devices, thermal management systems, and aerospace structures.

3. Our work on the automated optical micromanipulation system significantly reduces the time needed to conduct experiments and minimize damage to biological cells. This work has enabled biophysicists to conduct new experiments on interplay of cell-cell and cell-substrate adhesion in collective cell migration and make new discoveries.

4. We developed a new approach to utilize adaptive projection patterns in structured light-based 3D shape measurement. Our work demonstrated that patterns that use curved fringes with spatial pitch variation can significantly improve the measurement accuracy and coverage for parts that have highly curved geometries. This technology was commercialized by a leading metrology company.

5. We have developed Robo Raven, the first robotic bird capable of flying outdoors using independent wing control and performing aerobatic maneuvers. This platform is enabling new applications for aerial vehicles in surveillance and environmental monitoring applications.

6. We have developed methodologies for automated generation of alerts and re-tasking suggestions to assist human supervisors of multi-robot teams. This work improved the system performance by enabling human supervisors to make better decisions. Researchers studying human-machine integration at the Army Research Laboratory are using insights gained from this work.

7. We developed a manufacturing cell for enabling humans and robots to collaborate on performing composite prepreg sheet layup. This cell uses advanced machine learning for defect detection and digital twins for cell monitoring. In collaboration with multiple aerospace companies, we have demonstrated that robotic assistants can perform composite sheet layup tasks at human-competitive speeds and reduce human effort in ergonomically challenging tasks.

8. Traditional additive manufacturing (AM) capability is restricted to planar layered printing. We developed methods for using nonplanar material deposition in AM. This work has demonstrated the ability to create composite parts with the desired fiber orientation and led to dramatic increase in the performance of 3D printed composite parts.

9. We developed physics-informed AI to power smart robotic assistants that self-program and utilize sensor data to adapt and deliver efficient and safe operational performance in high-mix sanding and polishing applications. GrayMatter Robotics has commercialized robotic sanding and polishing technology for high-mix manufacturing applications.

10. We have developed a cell for enabling effective human robot collaboration for safely and efficiently performing assembly operations by leveraging the complementary strengths of humans and robots. We have developed methods to generate efficient plans by proactively accounting for the possibility of contingency events. We have demonstrated the usefulness of this approach in improving human productivity in aerospace assembly applications.