Robots have long struggled to move with the fluid grace of a human hand. While modern artificial intelligence can now write poetry, diagnose diseases, and generate stunning images, giving a machine the ability to physically manipulate the world remains a stubborn challenge. The core of the problem lies in how computers understand movement. Unlike text, which is made of distinct letters, or images, which are made of pixels, the physical actions of a robot arm are continuous streams of data. A robot arm does not simply jump from one position to another; it flows through an infinite number of tiny positions in between. To teach a robot using the powerful language models that drive today's AI, engineers must first translate this smooth, continuous motion into a series of discrete steps, much like turning a flowing river into a string of individual beads. This translation process is known as tokenization. If the translation is too rough, the robot loses the fine details of its movement, resulting in jerky, imprecise actions that fail at delicate tasks. If the translation is too complex, the computer cannot process it fast enough to react in real time.

For years, researchers have tried to solve this translation problem, but existing methods often force a square peg into a round hole. Some approaches treat every movement as a simple list of fixed categories, which ignores the subtle timing and relationships between steps. Others try to compress the data by grouping movements together, but they often lose the specific details needed for precise control, such as the exact angle of a wrist or the gentle squeeze of a gripper. This loss of detail creates a bottleneck, preventing robots from learning complex skills like stacking fragile objects or assembling intricate parts. The result is a robot that can perform broad, simple tasks but stumbles when faced with the nuanced demands of the real world.

A team of researchers has now proposed a new way to handle this translation, offering a method that preserves the richness of movement while keeping the data compact enough for modern AI to process. They call their system M2Tok, a tool designed to break down the continuous stream of robot actions into a format that language models can understand without losing the fine-grained dynamics required for success. Instead of treating the robot's entire body as a single, monolithic block of data, their approach splits the movement into separate, specialized channels. Imagine a robot arm that needs to move its shoulder, elbow, wrist, and gripper all at once. Older methods would try to compress all these different motions into one single code, often forcing the system to choose between accuracy for the arm and accuracy for the gripper. The new method, however, separates these motions into distinct groups, allowing the AI to focus on the specific nuances of each part independently.

The researchers achieved this by dividing the robot's movement data into multiple heads, where each head is responsible for a different aspect of the motion. One head might track the position of the arm, while another tracks the opening and closing of the gripper. Crucially, each of these heads uses its own independent dictionary of movement codes. By using multiple dictionaries working in parallel, the system creates a vast number of possible combinations, far more than a single dictionary could ever offer. This combinatorial power allows the system to represent a much wider variety of movements with high precision. It is similar to how a musician can create an infinite number of melodies by combining a limited set of notes across different instruments; the new system combines limited sets of movement codes across different "instruments" of the robot's body to recreate complex actions with remarkable fidelity.

To test this idea, the researchers trained their system on a large collection of simulated robot tasks, ranging from hammering a block to picking up diverse bottles and placing containers on plates. They compared their new method against several existing techniques that have been used in the field. The results showed a clear advantage for the new approach. In a series of twelve different simulation tasks, the robot using the new method succeeded in 51 percent of its attempts, significantly outperforming the next best method, which succeeded in only 45 percent of cases. The improvement was even more dramatic in difficult tasks that required high precision. For instance, in a task involving moving a can to a pot, the new method succeeded nearly twice as often as the previous best approach. In another task involving placing a burger and fries box onto a tray, the new method achieved a success rate of 64 percent, while the previous best method managed only 52 percent.

The researchers also tested the system on a different set of simulated environments to see if the skills could transfer to new situations. Here, the new method again demonstrated superior performance, achieving a success rate of 28 percent compared to 21 percent for the leading alternative. The most striking difference appeared in a task requiring the robot to pick up an eggplant and place it in a basket. The eggplant is an irregularly shaped object that is difficult to grasp, and previous methods failed to solve this task entirely. The new method, however, succeeded 33 percent of the time, suggesting that its ability to capture fine details allowed it to handle the complex geometry of the object where others could not.

To ensure these results were not just a product of the simulation, the team took their trained models and tested them on real-world robots. They used a mobile platform equipped with four robotic arms and a camera, asking the robots to perform three different tasks: ringing a bell, placing a container on a plate, and picking up different bottles. These were zero-shot tests, meaning the robots had never seen these specific real-world objects or environments before; they had to rely entirely on what they learned in the simulation. The new method again outperformed the others, achieving an average success rate of 33 percent across the three tasks, compared to a maximum of 28 percent for the best alternative. The visual evidence from these tests showed the robots successfully identifying the positions of objects and executing precise grasping operations, confirming that the high-quality translation of movement data held up when the robots left the computer and entered the physical world.

Beyond accuracy, the researchers also looked at how fast the system could run. For a robot to be useful, it must make decisions quickly enough to react to its environment. The new method allowed the robot to operate at a frequency of over 8 hertz, meaning it could make more than eight decisions every second. This is a significant improvement over older methods that operated at just 2 hertz. Furthermore, when the researchers optimized the system for speed using a specialized processing engine, the robot could reach a frequency of 56 hertz, far exceeding the speed required for most real-time manipulation tasks. This combination of high precision and high speed suggests that the new method could be a practical solution for deploying advanced robots in dynamic environments.

The success of this work hinges on a fundamental shift in how movement data is processed. By rejecting the idea that all movement must be compressed into a single, uniform code, the researchers allowed the system to respect the unique nature of different parts of the robot's body. The separation of movement into independent channels, combined with the use of multiple specialized dictionaries, created a system that could capture the subtle, high-frequency details of motion that previous methods missed. This approach did not require changing the underlying architecture of the language models that drive the robots; instead, it provided a better way to feed the data into them. The result is a robot that can understand and execute complex physical tasks with a level of dexterity that was previously out of reach, bridging the gap between the digital intelligence of language models and the physical reality of the world they are meant to navigate.