Social Humanoid gives a Unitree G1 a way to produce coordinated, whole-body responses to a nearby person, rather than simply replaying isolated motions. It combines a 260-hour human–humanoid interaction benchmark, a method for adapting human-human interaction clips to robot bodies, and a generative model that runs on the robot; demonstrations in simulation and on a physical G1 show responses aligned with the human partner.
What did the researchers build?
Social Humanoid is a system for generating whole-body movement on a Unitree G1 in response to a human partner. Its goal is social, physical interaction: motion where a robot and person coordinate their positions and may make contact, rather than a robot merely recognizing a gesture or moving independently.
The system combines three main pieces. First, an Adaptive-HHI Retargeting method converts human-human interaction recordings into paired human–humanoid motion. “Retargeting” means translating movement from one body to another with different proportions and joints. Second, a shared-codebook VQ-VAE converts movements from both bodies into a common set of discrete motion units. Third, a Diffusion Forcing Transformer uses that representation to generate a humanoid response conditioned on the person’s movement. The generated motion is executed on the G1 through SONIC.
The researchers also describe real-time shadow teleoperation, in which a human operator’s movement can drive the humanoid through the shared motion representation. Together, the system and its 260-hour HHI Benchmark address a basic development problem: social humanoids need examples of interactions between people and robots, but those recordings are scarce compared with human-human motion data.
What results did Social Humanoid achieve?
The reported result is a working pipeline that generates coordinated, human-aligned responses in simulation and on a physical Unitree G1. The researchers report that it handles a range of interaction types, and they present the 260-hour HHI Benchmark as a resource for training and evaluating social humanoid systems. The benchmark’s scale is notable because interaction data involving robots is a key bottleneck for this kind of work.
The approach also tackles a specific failure in straightforward motion transfer: contact points can shift when a human’s pose is copied onto a robot with very different proportions. The G1 is shorter than the source human, for example, so a partner’s hands could end up above the robot’s shoulders during a hug. The researchers’ two-pass retargeting process adapts the partner’s pose as well as the robot’s, aiming to preserve contact and the geometry between them.
The available results do not provide numerical success rates, latency measurements, or head-to-head performance gains over earlier systems. There is therefore no basis here for ranking Social Humanoid against other methods by a single score. Its demonstrated contribution is the combination of interaction-preserving data generation, a shared motion representation, and real-robot execution.
How does the system work?
The pipeline starts with human-human interaction clips. Directly mapping one person’s joint positions onto a humanoid is unreliable because the bodies differ in height, limb length, and joint structure. The retargeting method is designed to keep the interaction itself intact: where the participants touch, how their poses relate, and how they are positioned relative to each other.
Retargeting runs in two passes. In the first, a human’s motion is adapted to the humanoid while the other person remains unchanged. That can leave contacts misaligned: a partner may reach too high for the shorter G1. In the second pass, the partner’s pose is adjusted to match the retargeted robot. During frames with arm contact, inverse kinematics—a method for finding joint angles that place a limb at a target position—adjusts the partner’s shoulder, elbow, and wrist. For segments with substantial contact, a body-pose optimization also adjusts key body areas, including the wrists, chest, neck, hips, spine, and knees, while respecting human joint limits.
Each clip is retargeted again with the two interaction roles swapped, so the data represents both sides of the exchange. The resulting human and humanoid movements are encoded by a shared-codebook VQ-VAE. In plain terms, the model represents both bodies using a common vocabulary of learned motion units. A Diffusion Forcing Transformer then generates humanoid motion in that shared space, conditioned on the human partner. SONIC carries the generated motion onto the robot; the same shared representation also supports shadow teleoperation.
Why does this matter for robotics?
A socially responsive robot has to do more than detect that a person is present. It needs to move in ways that fit the interaction: approaching at an appropriate distance, coordinating its posture, and maintaining a shared position during contact. Those capabilities matter in settings where robots work near people, including assistance, collaborative tasks, and demonstrations. This research provides a route for teaching those behaviors from recorded human interactions instead of relying only on manually scripted motions.
The key engineering insight is that social behavior depends on the relationship between two bodies. A robot can produce a plausible-looking pose and still reach the wrong place or break contact because its proportions differ from a human’s. Adapting both the robot and partner motions creates training examples that better reflect the geometry of a real exchange.
The work is also relevant to humanoid developers evaluating platforms for interaction-focused applications. The humanoid robot marketplace on Robot Overflow provides a place to explore available humanoid systems. Social Humanoid is a research system, not evidence that a G1 is ready for unsupervised public interaction; its value is in showing how data generation and motion generation can be connected on a physical platform.
What are the limitations and open questions?
The system depends on retargeted recordings, and its contact-preserving method has to handle substantial differences in body shape without producing unnatural human poses. The authors describe joint limits and body-pose adjustments, but the reported material does not quantify how often retargeting succeeds across different interactions or body types.
The real-robot demonstrations establish that the approach can run on a G1, but they do not by themselves show safe, reliable operation across long deployments, varied environments, or unpredictable partners. Physical contact also raises questions about force control, collision avoidance, and how the robot should respond when an interaction changes suddenly. Further evaluation would need clear measures of motion quality, contact accuracy, response timing, safety, and human comfort. Numerical comparisons with previous systems would also help clarify where the method offers the largest gains.
What questions are frequently asked about Social Humanoid?
What robot does Social Humanoid control? The system generates and executes whole-body motion on a Unitree G1 humanoid.
What is the 260-hour HHI Benchmark? It is a human–humanoid interaction benchmark introduced alongside the system. The reported summary identifies its duration but does not detail its full recording composition.
How does the method handle the robot’s different body proportions? It retargets the robot and then adjusts the human partner’s pose to restore contact and interaction geometry while respecting joint limits.
Does the system work on a physical robot? Yes. The researchers report demonstrations in simulation and on a real G1, though the reported material does not establish performance in broad, unsupervised deployments.
What is the bottom line?
Social Humanoid connects interaction-aware motion data, a shared representation for people and robots, and generative whole-body control on a real humanoid. Its central contribution is a practical way to preserve contact and relative positioning when human interactions are adapted to a robot’s different body.