The Potential of Robotic Manipulation with Professor Russ Tedrake

Written by Audrey Woods

Like all specialties in computer science, the field of robotics is almost unrecognizably different from when it began. One hundred years ago, when the word “robot” was first coined in a Czech play, the idea of machines that could do automated tasks and help people in everyday situations was reserved for fiction and futurists. But now robots are everywhere, assembling in factories, assisting in surgeries, and cleaning the floors of over 14 million American households in the case of the robotic vacuum.

However, CSAIL Professor Russ Tedrake believes that the field of robotics is on the verge of a massive transformation, with “dramatically more capable robots” soon to come. As the Toyota Professor of Electrical Engineering and Computer Science, Aeronautics and Astronautics, and Mechanical Engineering at MIT and the Director of the Center for Robotics at CSAIL, Professor Tedrake is ready to be a part of what he believes will be an “incredibly bright” future for robotics.

Finding His Interest

Professor Tedrake has always loved robots and wanted to work with them since before he began his undergraduate studies. Originally, he planned to study mechanical engineering like his father, but his career trajectory changed when he took his first programming course as a freshman and got “hooked on computer science.” This interest led him to major in computer engineering at the University of Michigan and then come to MIT for his doctorate, where he wrote his PhD on reinforcement learning.

While at MIT, Professor Tedrake developed a passion for robotic control theory in part after learning about passive dynamic walkers at the MIT Leg Laboratory. Passive dynamic walkers are, as Professor Tedrake describes, “incredibly elegant walking machines that move gracefully down a small slope powered only by gravity.” He says, “the contrast from the way these machines moved vs the way our ‘controlled’ robots moved was very dramatic and very motivating for me.” After earning his degree and joining the MIT faculty, Professor Tedrake went on to collaborate with other professors at MIT, especially EECS Professor Alexandre Megretski and EECS Professor Pablo Parrilo, whose work framed understanding very complex systems through the lens of optimization and control. These ideas have played a major role in Professor Tedrake’s research since then.  

Despite his early passion for applying learning concepts in robotics, Professor Tedrake says he was “too early” with the reinforcement learning ideas he explored in his graduate thesis. In his course notes for Underactuated Robots, Professor Tedrake uses a story of when he was a young faculty at a robotics conference to illustrate how differently the field used to think about robotic intelligence. When talking to other experts at the event, Professor Tedrake says, “One of the senior faculty said ‘Russ: the people that talk like you aren't the people that get real robots to work.’” Now the field has come to embrace using optimization or learning in robotic planning and control systems so much that Professor Tedrake jokes he has to defend himself if he doesn’t use learning in a robotics project.

Nowadays, Professor Tedrake is looking over the horizon of another major shift in the field with his work on robotic manipulation. He believes this area of research is “about to have our break-through… with a general control strategy comparable to Chat GPT but for robot dexterity.” If true, this would make robots more versatile, useful, and widely applicable, not to mention change the way engineers approach control theory. For this reason, Professor Tedrake is “intensely interested in this and pursuing it aggressively.”

Robotic Manipulation, Foundational Models, and Training Future Roboticists

One of the avenues through which Professor Tedrake is exploring robot dexterity is in his role as the Vice President of Robotics Research at the Toyota Research Institute (TRI). At TRI, for example, researchers are creating robots like Punyo that are equipped with sensors, padding, and soft grippers that allow it to use whole-body manipulation techniques. As with humans, empowering the robot to use its torso and arms in lifting and manipulating allows it to handle heavier, bulkier objects more efficiently. The ethos behind Punyo is to create safe robots that amplify what people can do and assist in everyday activities, like carrying groceries or putting away laundry. Punyo can be trained using either human demonstration or high fidelity simulation, which gives the robot a range of ways to learn new skills.

Professor Tedrake’s team at TRI is also applying diffusion policy in robot learning, which allows researchers to teach robots faster and with fewer demonstrations. This is the basis of what he calls “Large Behavior Models,” which, like Large Language Models, would create a foundation of skill and competence that could be widely applied and rapidly adapted. During a robotics panel at the 2024 CSAIL Alliances Annual Meeting, Professor Tedrake explained, “The new goal is to take the foundation models we’ve seen in other disciplines and bring them into robotics.” According to Professor Tedrake, the TRI’s “kindergarten for robots” will teach thousands of new skills using demonstration, simulation, and fleet learning, which connects robots and allows a skill learned by one machine to be transferrable to another. While right now the TRI robots are still being trained skill by skill, Professor Tedrake says, “the next breakthrough will be when we’ve trained the robots with enough dexterous skills that they’ll be able to generalize, performing a new skill that they’ve never been taught.”

When deploying robots in the real world—as Professor Tedrake hopes to do—optimization becomes an important and challenging consideration. The current method for optimization is sampling-based planners, but these don’t scale well and struggle with continuous differential restraints. To address this, Professor Tedrake’s research group created an optimization framework called Graph of Convex Sets, which, when it comes to robotics, “can find better trajectories in less time than widely used sampling-based algorithms and can reliably design trajectories in high-dimensional complex environments.” Much of the work was funded by Amazon Robotics, who challenged Professor Tedrake’s team to design methods that would enable their robot arms to move at high speeds despite working in complex and cluttered environments.

In addition to helping bring about the future of robotics, Professor Tedrake is also playing a key role in educating and training future roboticists. His passion for robotic manipulation fuels his EECS course on the subject, which he developed to teach students the fundamentals of designing robots that can autonomously manipulate objects in unstructured environments. The course invites students from a variety of backgrounds to experiment with robotic manipulation and explore creative solutions in this rapidly evolving field. Assisting in many course projects is Professor Tedrake’s Drake program, an open-source toolbox “for analyzing the dynamics of our robots and building control systems for them, with a heavy emphasis on optimization-based design/analysis.” In other words, Drake makes robotics more accessible and applicable at scale. The approachability of both Drake and the Robotic Manipulation course has meant that even industry professionals and non-students are using Professor Tedrake’s teaching materials (such as the Robotic Manipulation course notes) to imagine and design the robots of tomorrow.

The Future of Robotics

While the idea of robots that can assist in complex, dexterous activities or be deployed in dynamic environments might seem far away, Professor Tedrake says, “we're on the verge of having dramatically more capable robots that can operate in ‘open worlds,’ tasked by natural language specifications.” This will lead to robots that integrate into our lives and assist in daily activities such as cooking, carrying, sorting, and more.

To reach this future, Professor Tedrake believes it’s important to scale our robot learning and build the foundational models that will enable these multi-skilled and versatile machines. Part of that will happen in industry, where many such robots will be built, tested, and sold. But Professor Tedrake emphasizes that the academic focus on theory will be equally important. “It's also essential that we continue to develop our theoretical understanding of these tools, and develop the next wave of algorithmic advances,” he says.

Generally, Professor Tedrake feels that “this is an incredible time to be a roboticist.”  

Learn more about Professor Tedrake on his website or CSAIL page