PAMoR: Parameterized Affective Motion Generation in Real Time for Humanoid Robots

In the crowded spaces of our daily lives, from shopping malls to concert stages, robots are beginning to move among us. We do not just watch them to see if they can lift a box or walk across a room; we watch them to see how they feel. Humans are wired to read emotion in movement. A slumped shoulder suggests sadness, while a bounding step implies joy. This is a language of the body that we understand without words, a skill we use constantly when interacting with other people. But for a robot to speak this language, it must do more than just perform a task; it must perform that task with a specific emotional tone. Until now, teaching a machine to do this has been difficult. Most systems that create emotional movement for digital characters rely on human actors to provide a reference video or a simple label like "happy." These methods are rigid. A label cannot be tweaked to make a movement slightly more excited or slightly less tense, and a video reference is locked in time, unable to change its mood on the fly. Furthermore, when these digital movements are transferred to a physical robot, the process often distorts the very feelings they were meant to convey, losing the subtle shifts in speed and posture that carry the emotion.

A team of researchers has developed a new way to solve this problem, creating a system that allows a physical robot to generate emotional movement in real time, with the ability to adjust its mood as it moves. They call their framework PAMoR. Instead of relying on human labels or pre-recorded videos, the system treats emotion as a set of measurable physical quantities. It uses a psychological model known as the valence-arousal plane, which maps feelings onto two simple scales. The first scale, valence, measures how positive or negative a feeling is, ranging from sadness to happiness. The second scale, arousal, measures how active or calm a feeling is, ranging from boredom to excitement. The researchers discovered that they could calculate these two numbers directly from the robot's own body movements. They found that a positive, happy feeling corresponds to an open, expanded posture, while a negative feeling corresponds to a contracted, folded one. Similarly, high energy and speed indicate high arousal, while slow, still movements indicate low arousal. By turning these physical observations into a mathematical formula, the system can label every movement it sees with a precise emotional coordinate, without needing a human to watch and write it down.

The researchers built a robot capable of using this system on a 29-joint humanoid machine called the Unitree G1. The robot's brain is designed to listen to three different instructions at once. One instruction tells it what action to perform, such as walking or punching. The other two instructions tell it how to perform that action by setting the valence and arousal numbers. To make this work without the instructions getting confused with one another, the system uses a special method where it combines three separate learning models. One model learns the action, while the other two learn how to shape that action with different emotional tones. At every step of the movement generation, these models work together to produce a motion that fits the command. This allows the robot to change its mood while it is moving. A user can tell the robot to walk, and then, while it is walking, change the command from "calm and neutral" to "excited and happy," and the robot will instantly adjust its stride and posture to match the new feeling.

The results of this work show that the system works remarkably well. When the researchers commanded the robot to move with specific emotional coordinates, the robot's actual movements matched those commands with high precision. The system could generate a full range of emotions, from distressed to relaxed, and the robot's body language shifted smoothly across this spectrum. To see if humans could actually read these emotions, the researchers conducted a study where people watched videos of the robot performing actions like waving or punching with different emotional tones. The observers were asked to guess what emotion the robot was expressing. The system succeeded in making its intended emotions clear to the viewers. In nearly 40 percent of the trials, the viewers correctly identified the specific emotion the robot was trying to convey. This is a significant improvement over previous methods, which relied on text prompts or video references, and it approaches the level of recognition people achieve when watching human actors perform emotional movements. The study also confirmed that the robot's movements felt natural and human-like, rather than mechanical or stiff.

This approach represents a shift in how we think about robot movement. Instead of treating emotion as a separate layer added on top of a task, the researchers grounded it in the physical reality of the robot's body. By measuring posture and speed directly, they created a system where emotion is not just a label, but a controllable parameter that can be adjusted with the same ease as volume or speed. The system operates in real time, meaning the robot can react to its environment and change its emotional expression instantly. While the researchers note that the system currently relies on a pre-programmed understanding of how body shape relates to emotion, and that future work will need to address how the robot maintains this expression when its physical controller makes small errors, the foundation is solid. They have demonstrated that a robot can generate complex, emotional whole-body motion on the fly, and that humans can read those emotions with surprising accuracy. This brings us closer to a future where robots are not just functional tools, but social partners capable of expressing themselves in the universal language of the body.