Brown University Pioneers AI-Powered Humanoid Motion
Inspired by the futuristic world of The Jetsons, researchers at Brown University are working to bring humanoid robots into reality. While flying cars remain elusive, humanoid robotics is rapidly advancing, largely thanks to breakthroughs in artificial intelligence. At the forefront of this innovation is Srinath Sridhar, assistant professor of computer science at Brown, who, along with his team, has developed a new AI model that allows robots to perform movements based on simple text commands.
MotionGlot: A Unified Approach to Robot Movement
Sridhar and Sudarshan Harithas, a Ph.D. student in computer science, collaborated with a team of graduate and undergraduate students to develop MotionGlot. This model aims to simplify how different kinds of robots are trained to move. Traditionally, each robot required a unique control model, but MotionGlot introduces a universal framework that uses English as an intermediary language to translate movement commands between robots.
“The goal is to create a single model that can work across various robots,” Sridhar explained. “We don’t want to develop a new system for every new type of robot. Instead, MotionGlot generalizes across platforms.”
The Role of AI in Humanoid Robotics
According to Sridhar, developing the physical hardware for humanoid robots is only part of the challenge. The real key lies in building a sophisticated AI “brain” that can interpret the world and make intelligent decisions. While current models handle basic walking and running, more intricate tasks—especially those involving the hands—remain unsolved.
“What’s missing is dexterity,” Sridhar said. “Most humanoids can walk, but manipulating objects with their hands is still a massive hurdle.”
Learning from Mistakes: AIDOL’s Fall
Sridhar referenced Russia’s humanoid robot AIDOL, which stumbled during a demonstration. He believes the issue wasn’t mechanical but rather the software controlling the robot. “It shows that without advanced AI models, even well-built robots can fail,” he noted. “The U.S. and China are currently leading in this space.”
Inspired by Large Language Models
Sridhar drew inspiration from the success of large language models like ChatGPT. “These models are incredibly effective at generating text, and we thought—why not apply similar principles to motion?” he said. MotionGlot doesn’t create a new programming language but instead uses existing robotic parameters and translates them through English to adapt between different robots.
“Each robot has its own set of parameters, such as X-Y-Z coordinates for limbs,” Sridhar explained. “MotionGlot serves as a translator, making the transition from one robot’s motion to another seamless.”
Next-Token Prediction in Robotics
One of the core technologies behind MotionGlot is “next-token prediction,” a method commonly used in natural language processing. The model takes a sequence—such as motion data or text—and predicts what comes next. This helps the AI anticipate movements even in unfamiliar scenarios.
“We train the model with thousands of motion pairs, including human, humanoid, and quadruped movements,” said Sridhar. “As a result, it can generalize and respond to commands it has never seen before.”
Real-World Applications and Embodiment Learning
The ultimate goal for Sridhar’s team is to create a universal model adaptable to any robotic platform. This approach, known as embodiment learning, is a growing field with multiple billion-dollar startups involved. “Many companies are chasing this vision,” Sridhar said. “What makes our work unique is the use of language translation concepts in robotics, which is still relatively new.”
Next Challenges: Robotic Hands
While MotionGlot demonstrates promising results with robot bodies, the next frontier is robotic hands. Sridhar emphasized that humans rely heavily on their hands for dexterous tasks, and teaching robots to replicate these movements is a major focus moving forward.
“We’re not just interested in translation tools,” he said. “We want to create fundamental learning methods so robots can observe and mimic human hand activity.”
Timeline and Funding
Although progress is being made, Sridhar estimates that fluid, natural robot movements are still at least five years away. One major limitation is the lack of high-quality data. “We need better datasets to train these systems effectively,” he noted. “That’s a significant area of our current research.”
Funding from the Office of Naval Research (ONR) has played a crucial role in supporting this high-risk, high-reward research. “We’re grateful to ONR for backing not just this project but several others,” Sridhar said. “Their support allows us to explore bold ideas that could redefine the future of robotics.”
This article is inspired by content from Original Source. It has been rephrased for originality. Images are credited to the original source.