Written by Matthew Busekroos | Produced by Andrew Zucosky
Originally from San Diego, undergraduate Rohan Bosworth is studying 6-4 (Artificial Intelligence and Decision Making) at MIT. Bosworth is a UROP at CSAIL working with Professor Russ Tedrake and the lab’s Robot Locomotion Group (RLG).
During high school, Bosworth was part of the Existential Robotics Laboratory at UCSD, led by Professor Nikolay Atanasov. There, he developed a strong foundation in robotics systems, working on projects that applied state-of-the-art algorithms for mapping, localization, and path planning. Over time, Bosworth gravitated toward refining graph-based motion planning algorithms, leading to his first publication at the International Conference on Intelligent Robots and Systems.
Bosworth also competed in the FIRST Tech Challenge from seventh grade through high school, which sparked his passion for designing and implementing real-world systems. When he began searching for UROP opportunities, Bosworth said CSAIL stood out not only for its reputation at the forefront of research but also for its emphasis on building practical, impactful systems.
Bosworth noted the Robot Locomotion Group, in particular, inspired him with its commitment to bridging the gap between academia and industry. He said the lab’s collaborations with leading companies directly apply research innovations to real-world problems, and he was fortunate to secure a summer internship at Sony through one of these collaborations.
“Working with Prof. Tedrake and the RLG has been an unparalleled experience,” Bosworth said.
“When I joined the lab as a first-year student, I was unfamiliar with many of the cutting-edge topics being explored in the group. However, the mentorship I received from Prof. Tedrake and the graduate students in the lab was beyond anything I could have imagined.”
Bosworth said his first project was designing a clutter-clearing demonstration using one of the lab’s key research innovations: tools for planning shortest paths on Graphs of Convex Sets (GCS). This framework, which combines traditional graph search algorithms with advanced convex optimization techniques, enables fast collision-free robot motion planning. Despite Bosworth’s initial unfamiliarity with mathematical programming, he quickly developed the required skills thanks to the guidance of his mentors.
“Through this experience, I learned more than I could have expected, from robot manipulation to diffusion policies,” he said. “I explored the application of optimization to robotic systems, particularly complex systems like robotic arms, and worked directly with physical robot platforms. This hands-on experience allowed me to confront the challenges of real-world robotics and provided the foundation for my current research into applying diffusion policies to manipulation tasks.”
Bosworth’s research has evolved significantly over the past year and a half, and he is currently focused on applying diffusion policies to robotics. His initial project was to apply Graphs of Convex Sets (GCS) to a clutter-clearing demonstration, relocating scattered objects from a bin.
GCS is a novel algorithm combining traditional graph theory algorithms and advanced optimization techniques into one elegant mathematical programming framework. GCS can be applied to a large range of problems from shortest trajectory path planning to large scale task planning. According to Bosworth, the ultimate goal of this project was to evaluate GCS and discover pain points as well as to demonstrate how GCS could be applied to real world systems.
“In the final approach, I utilized camera systems to evaluate where a 7-joint robot arm should grasp an object from a bin and then used GCS to pick up an object from one bin and relocate it to another,” he said. “This is then repeated until all the clutter from the original bin was relocated to the other. While seemingly simple, this task illustrated many critical challenges especially the one of state estimation or perceiving the state of your environment.”
In the application of the clutter clearing problem, every time objects are added to the pickup bin or an object is removed from the bin, the environment changes, Bosworth noted. This constantly evolving environment is a critical struggle and led to many problems in colliding with other objects within the bin environment as they were not properly perceived or struggling to account for how robot movement must change when it is carrying a grasped object. A critical step in planning and control, state estimation often introduces errors, particularly when perceiving an environment with noisy measurements and applying heuristics to decisions such as grasping.
“For example, one area I struggled with was camera calibration or determining the relative pose of a camera system,” he said. “There were limited existing packages for calibration available for our platform and when my collaborators utilized them they often found the calibrations to have error that would inhibit the ability to accurately decide the location of objects in the environment. Camera calibration was one of many problems I encountered which made optimization techniques a challenge to work with specifically for this problem. However, it is important to note, optimization techniques were wonderful at solving certain problems.
For example, following a grasp of an object, an optimization technique could provide a full trajectory to a deposit point from most object grasp points in seconds, something that few methods could achieve. Moreover, optimization is able to provide a quality of trajectories that resembles human like movement. For certain problems such as navigating an evolving and random environment such as a bin full of objects, GCS based trajectory planning may not be the optimal tool, but there are certainly instances where GCS is likely the best solution to the motion planning problem.”
To address these limitations, Bosworth is exploring learning-based techniques. Traditional learning approaches, such as imitation learning and reinforcement learning, typically rely on human demonstrations or extensive exploration. Bosworth’s approach investigates whether a learning-based system can leverage data from model-based methods instead.
Specifically, Bosworth is working to implement diffusion policies—an emerging concept inspired by the success of generative AI in domains like image generation. Diffusion policies function by taking a set of noisy actions or actions that a robot could take that are supposedly random and then estimating the noise in these actions conditioned on environmental information. This noise is then estimated and removed from the actions repeatedly until they are no longer noisy or they represent the actions the robot should take in this case.
Diffusion policies have been shown to have high performance with limited amounts of data and also encapsulate multi-modal situations where traditional models may struggle, according to Bosworth. These methods have shown promise when applied to robotic action generation, offering a novel avenue for more robust and adaptive decision-making systems. Bosworth seeks to evaluate how learning and model based methods could be blended to allow for a stronger overall system or even a continual learning platform. He said he believes the strong areas of these two methods are distinct which means that they can be meshed to have even stronger results.
“Robotics has the extraordinary potential to bridge the digital and physical worlds,” Bosworth said. “My long-term goal is to design systems that allow robots to seamlessly perceive, decide, and act in complex, dynamic environments, enhancing their utility across a wide range of applications.”
Bosworth said his passion for robotics stems from a decade of involvement in competitive robotics and a fascination with the mathematical underpinnings of autonomy and decision-making. He added that he is deeply inspired by the potential of robotics to transform human well-being.
“Robots have the potential to transform human well-being by addressing critical challenges across various domains,” he said. “For example, autonomous robots can support search and rescue operations in hazardous environments, navigating rubble or unstable terrain to locate survivors. Similarly, robotic systems can deliver essential medical care in underserved regions, enabling remote surgeries or transporting lifesaving supplies to inaccessible areas.”
Bosworth said among the many challenges in robotics, he is particularly drawn to the problem of robot decision-making—the essential process that determines how robots interact with the world. He added that it is a fundamental, yet deeply complex, problem that he finds both intellectually and practically rewarding to tackle.
After completing his undergraduate studies, Bosworth plans to pursue a Masters of Engineering and PhD in robotics. Admittedly, Bosworth said he was initially uncertain about pursuing graduate studies. However, his experience with the RLG and the supportive culture of CSAIL convinced him to take this path.
“The collaborative, innovation-driven environment of CSAIL has been instrumental in shaping my aspirations, and I am eager to contribute to the field through advanced research and mentorship in the future,” Bosworth said. “This alignment of cutting-edge research and real-world impact was exactly what I was looking for. Moreover, the RLG has an approach to research that is quite unique. Throughout the lab there is a focus on careful thinking and the application of rigorous mathematical tools to robotics and beyond. While I was not aware of this culture prior to joining the group, I am quite grateful for it now. By thinking about problems at the fundamental level, I find that the group is able to draw truly insightful conclusions. As such, I am hopeful to carry this culture forward in my future endeavors.”