Autonomous robots have demonstrated enormous potential in fielded environments; however, to be broadly accepted, they need to demonstrate safe and resilient performance in a wide range of varying circumstances, from degraded environmental conditions to different terrains and adverse obstacles. To achieve scalability, these techniques need to be universal. This unconstrained nature of robots in fielded environments makes hand design of reliable systems impossible and defining requirements difficult. My research goal is to establish a new research area called “resilient robotics” that addresses these challenges.
The key to autonomous robots in highly uncertain scenarios is to exhibit safe and resilient performance [1]. While highly performant systems have been engineered, they are often brittle or need to be carefully handcrafted to achieve sufficient capabilities. On the other hand, provably safe systems typically have insufficient performance.
Over the last decade in my lab, I have achieved resilient performance of robots by advancing the robustness, redundancy, and resourcefulness of the algorithms as well as systems. While careful engineering is part of resilient performance, I am particularly interested in answering the following fundamental questions:
- How does one design algorithms and systems that are robust in the face of large uncertainty and learning-based components?
- Where can we inject redundancy into the system without incurring excessive computation or weight penalties?
- How can we move beyond fixed behaviors, policies, or interpretations of the data and have a continuous improvement of our systems to achieve resiliency in the face of large uncertainty with little data?
These research questions address fundamental limitations of today’s autonomous systems, where the complex interactions between different modules lead to fragility and unexpected behaviors, while the desired behavior is difficult to capture. This has led to a new line of inquiry into “resilient robotics.”
I have been studying these questions by researching and applying machine learning methods that continuously improve the system with more data, with an emphasis on explainability, to the traditional perception/state estimation/and planning components of autonomy that rely on strong models as well as blending the approach with model-free methods where appropriate.
In my research, I use the terms model and adaptation broadly: Models can be robot dynamics, error dynamics of state estimation, wind model, planning abstractions, or perception representations, for example. Adaptation based on “error signals” could be an integrator, self-supervised learning, or offline training of neural models.
Current approaches to improving operations of autonomous vehicles fall short. Either they can reactively keep the vehicles safe, have high performance in the nominal cases, or be able to adapt but not perform relevant missions. These shortcomings are because either approach relies on static models and does not utilize the feedback available or only relies on training of policies without regard to safety.
I have found that adaptive model-based approaches are powerful and high-performance when I applied them to autonomous flight, offroad driving, and multi-agent systems. The resulting models can be powerful because they enable the abstraction of irrelevant parts, the ability to predict and reason based on these predictions. However, I have also found that models are fragile and will fail if the modeling error is too large or if the environment changes.
Any useful model will have some simplifying assumptions, and if these assumptions are valid, the overall system will work well. Modelling errors are typically easy to find; however, how to adapt to these errors is not obvious. On the other hand, model-free approaches have the advantage of not placing an artificial constraint on the complexity of the model or where model complexity needs to be expressed; however, the performance outside of the observed boundaries is unknown, and validating the correctness is difficult.
Over the last decade, I have made fundamental contributions to this new area of “resilient robotics” to answer those key research questions for SLAM, perception, and planning, by demonstrating pioneering results, as well as by evaluating the resilience in the context of applications such as subterranean exploration, search & rescue, triage, wildfire, safety in shared airspace, autonomous offroad driving, autonomous full-scale helicopter flight, bridge inspection, and drone delivery.
I explore resilient robotics by “grounding” research problems in impactful applications. My efforts are not limited by what is perceived as too difficult or too laborious. I embrace the challenge and build as necessary and leverage what already exists if possible. A large part of my effort goes into formulating what the core research problem is and then “cleaning up” these problems. Often, I find that existing problem formulations addressed in prior work have a fundamental gap in their assumptions to be able to be effective for relevant applications which require advancements in core methods. I strive to validate our methods in the field in closed-loop experiments, beyond benchmarking on datasets. I test early and test often since I have seen that these experiences lead to richer feedback for the systems, and as I gather more data, algorithms keep improving. It is now an exciting time since I can go beyond relying on smart engineering of solutions and can start making stronger assertions using large-scale evaluations.
Research Topics
current masters students
current affiliates
past masters students
- Vasu Agrawal
- Anish Bhattacharya
- Mateo Guaman Castro
- Po-Wei Chou
- Vishal Dugar
- Rohit Garg
- Sourish Ghosh
- Kristen Holtz
- Ratnesh Madaan
- Lucas Casanova Nogueira
- Kevin Pluckter
- Connor Pulling
- Aditya Rauniyar
- Andrew Saba
- Matthew Sivaprakasam
- Vaibhav (Vai) Viswanathan
- Luke Yoder
- Sam Zeng
- Hengrui (Henry) Zhang
Below is a list of this RI member's most recent, active or featured projects. To view archived projects, please visit the project archive