The Motionglot IA translates natural language into understandable movement controls and executable by robots.
In Providence, Rhode Island – researchers from Brown University have developed an AI model named Motionglot. This tool makes it possible to control the movement of robots and digital characters using simple textual instructions. Innovation simplifies the control of machines, a useful advance in robotics, animation and virtual reality.
The Motionglot translates a text in motion
A humanoid robot, a quadruped or an animated character do not move the same way. Even an action as simple as walking differs according to the morphology.
Until recently, the intelligent systems struggled to generalize an instruction to various types of body. The Motionglot IA takes up this challenge posed by the diversity of robotic bodies.
This model interpret an instruction as a translator Go from English to Chinese. Textual commands, such as ” move on and turn left “, Become a sequence of movements adapted to the physical structure of the robot or avatar.
To achieve this, the researchers adopted a Approach inspired by LLMsimilar to those who feed text generators. The Motionglot IA breaks down the movements into elementary units, called “ Movement tokens “, Comparable to words in a sentence.
By predicting the rest of these tokens, the model produces fluid and natural actions. For example, “walking happily” results in rhythmic steps for a humanoid or a lively trotting for a quadruped.
Training on various data
IA Motionglot training is based on two data sets. The first, quad-loco, contains Movements of annotated quadruped robots. The second, qua-cap, includes human movements associated with textual descriptions.
These data allow Motionglot to understand how the same intention varies according to the body. During the tests, this AI has demonstrated remarkable flexibility.
The model performs both specific orders (“back down, then turn”) and wave requests (“make cardio”), with responses adapted to the context.
Motionglot potential applications are vast, from robotics to video games, including animations. However, this tool has limits.
The model was tested on controlled data, thus requiring more information To be fully deployable on a large scale.
Naturally, researchers wish to make Motionglot accessible to the public. The project already benefits from the support of the Office of Naval Research. This research will be presented at the International Conference on Robotics and Automation 2025, from May 19 to 23 in Atlanta.
.