Human2Humanoid: Physics-Aware Cross-Morphology Motion Retargeting for Humanoid Robots

summary

The gist

Human2Humanoid presents an unsupervised motion retargeting framework that transfers human motions to humanoid robot behaviors with high fidelity This paper presents Human2Humanoid…

In short

Human2Humanoid is an unsupervised system that transfers human motions to humanoid robot behaviors with high fidelity, even when the human and robot have different physical appearances. It uses a CycleGAN and a skeleton-aware graph convolutional network to capture motion topology. The framework ensures realistic results by enforcing consistency between end-effector trajectories across different body types and adding physics constraints for stable contact patterns.

Key concepts

CycleGAN
This is an architecture used to learn mappings between two different data domains without needing paired examples. In this case, it helps the system translate human motion data into robot motion data even though the human and robot bodies look very different.
Skeleton-Aware Graph Convolutional Network (GCN)
This network is designed to understand the structure of a body's skeleton. It uses graph convolutions to aggregate information from neighboring joints, ensuring that the motion transfer respects the actual physical connections and topology of both human and robot bodies.
Morphology-Invariant End-Effector Consistency Loss
This loss function ensures that the end-effector positions (like hands or feet) are aligned relative to a standard pose (T-pose) for each body type, rather than matching absolute coordinates. This makes the motion transfer consistent regardless of how different the robot's morphology is from the human's.
Physics-Aware Feasibility Constraints
These are rules added during training that enforce physical realism. They penalize unrealistic behaviors like feet floating or exceeding joint limits, ensuring that the resulting robot motions are physically plausible and can actually be executed by a real robot.

Terminology used across episodes

This episode discusses

The paper

Human2Humanoid: Physics-Aware Cross-Morphology Motion Retargeting for Humanoid Robots · Read on arXiv

Tianchen Huang, Feiyang Yuan, Junchi Gu, Shurui Fang, Xiaohu Zhang, Yu Wang, Wei Gao, Shiwu Zhang

University of Science and Technology of China

Retargeting human motion to humanoid robots is critical for teleoperation, imitation learning and human-robot interaction. However, it remains challenging because of substantial morphological discrepancies between humans and robots, including differences in skeletal topology, limb proportions and degrees of freedom, as well as the scarcity of paired motion data. This paper presents Human2Humanoid, an unsupervised motion retargeting framework that transfers human motions to humanoid robot behaviors with high fidelity. To bridge the domain gap under unpaired data, we adopt a CycleGAN-based architecture equipped with a skeleton-aware graph convolutional network to capture topology-dependent motion features. To address cross-domain scale mismatches, we introduce a morphology-invariant end-effector consistency loss that aligns normalized end-effector trajectories to preserve motion semantics across embodiments. To improve physical plausibility and reduce contact artifacts, we impose explicit physics-aware feasibility constraints to encourage reproduction of the contact patterns in the source motion. Experimental results show that the proposed method successfully retargets human motion to the Unitree G1 humanoid robot without paired data, and outperforms existing methods in both downstream controllability and physical feasibility.

DOI: 10.1109/LRA.2026.3733563

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Human2Humanoid: Physics-Aware Cross-Morphology Motion Retargeting for Humanoid Robots".

Dev: Here's the gist of the paper. Human2Humanoid presents an unsupervised motion retargeting framework that transfers human motions to humanoid robot behaviors with high fidelity This paper presents Human2Humanoid, an unsupervised motion retargeting framework that transfers human motions to humanoid robot behaviors with high fidelity To bridge the domain gap under unpaired data, we adopt a CycleGAN-based architecture equipped with a skeleton-aware graph convolutional network to capture topology-dependent motion features To address crossdomain scale mismatches, we introduce a morphology-invariant end-effector consistency loss that aligns normalized end-effector trajectories to preserve motion semantics across embodiments To improve physical plausibility and reduce contact artifacts, we impose explicit physics-aware feasibility constraints to encourage reproduction of the contact patterns in the source motion Framework Architecture The framework comprises two generators GB→A and GA→B, as well as two discriminators DA and DB The Skeleton-Aware Generator utilizes SkeletonConv layers to explicitly leverage the skeletal topology defined by the adjacency matrix, where the feature aggregation involves not only itself but also its neighboring joints N (j), expressed as fout(j) = Xk∈N(j)∪ðj)Wk · fin(k) + b (1) Furthermore, topology-based pooling strategies are employed to compress high-dimensional skeletal features into a lowdimensional latent space, followed by unpooling operations to recover the target topology The Frame-wise Pose Discriminator contains two complementary branches: the joint-wise pose branch, denoted as Deach, predicts joint-wise realism scores using joint-specific linear heads and the full-body pose branch, denoted as Dall ∈ R(N·T)×1, flattens the fullbody latent pose feature and predicts a pose-level coherence score through a Multi-Layer Perceptron Semantic Preservation Mechanism To address morphological discrepancies and ensure physical plausibility, the framework integrates Skeleton-Aware Graph Convolutional Networks (GCNs) and introduces Morphology-Invariant EndEffector Consistency Loss alongside Physics-Aware Feasibility Constraints The Morphology-Invariant EndEffector Consistency Loss aligns end-effector trajectories in a morphology-invariant space defined relative to the rest pose (T-pose) rather than directly matching absolute coordinates This loss is calculated as LEE = E

FKk(qB, t) − FKk(qTposeB) / SB − FKk(ˆqA, t) − FKk(qTposeA): This ensures that the end-effector positions are aligned relative to the T-poses of each embodiment Physical Plausibility Enforcement To enhance deployability on real robots, explicit foot contact, foot height and joint limit constraints are introduced during training The Physics-Aware Feasibility Constraints include three components: 1) Foot Contact Constraint: This penalizes the target foot velocity as Lcon = PΣtPm∈F c(m)B(t)v(m)A(t)/SA 2) Foot Height Constraint: This applies an anti-floating hinge penalty on the target foot height as Lhei = PΣtPm∈F w(m)(t)-hrefA,mSA squared 3) Joint Limit Constraints: A joint limit loss is introduced to penalize the predicted robot joint angles qˆA falling outside the range

qmin, qmax: as Llim = ΣXtXjReLU(ˆqA,t,j − qmax,j) squared + ReLU(qmin,j − qˆA,t,j) squared Training Objective and Results

Paper summary: Rosa: So we're looking at this paper called Human2Humanoid: Physics-Aware Cross-Morphology Motion Retargeting for Humanoid Robots. Essentially, the authors are tackling how to take motion from a human and make a humanoid robot do it, even when the human and the robot look completely different.

Dev: Right. The big challenge they're facing is that there's this huge gap between humans and robots—things like different skeletal shapes or how many joints they have—and most of the time, we don't have paired motion data to train a robot specifically for every single human body type.

Taro: So, the thesis here is building an unsupervised framework that transfers human motions to humanoid robot behaviors with high fidelity without needing those specific paired datasets for every embodiment.

Rosa: Exactly. To do this under those unpaired conditions, they're using a CycleGAN-based architecture combined with something called a skeleton-aware graph convolutional network to capture the motion features based on the skeletal topology.

Dev: That graph convolutional network part is key because it explicitly looks at how the joints are connected, not just treating each joint in isolation. They use layers that aggregate information from neighboring joints as well as itself, like this equation: fout(j) equals Xk in N(j) union j Wk fin(k) plus b.

Taro: And they use topology-based pooling strategies to squash those complex skeletal features down into a smaller latent space, and then unpooling operations to bring the target skeleton back up. It’s about encoding the shape into a shared space so you can retarget across different body types.

Rosa: They also have this morphology-invariant end-effector consistency loss that tries to align the end-effector trajectories relative to a standard T-pose, which helps keep the motion semantics consistent even when the bodies are scaled differently.

Dev: That loss formula they introduce is LEE equals E of FKk(qB, t) minus FKk(qTposeB) over SB minus FKk(ˆqA, t) minus FKk(qTposeA). It’s forcing the end-effector positions to line up relative to each robot's T-pose instead of just matching absolute coordinates.

Taro: And for the physical side, they added explicit physics-aware feasibility constraints during training to make sure the resulting motion is actually believable when you put it on a real robot.

Rosa: They have three specific constraints for that: one for foot contact, one for foot height, and one for joint limits. These are meant to prevent things like foot skating or the robot floating off the ground.

Paper summary: Dev: The foot contact constraint penalizes the target foot velocity as Lcon equals P summed over time t of Pm in F c(m)B(t)v(m)A(t)/SA, which tries to manage how fast feet move relative to their contact points on the ground.

Taro: And then there's the foot height constraint, which is an anti-floating hinge penalty that applies a weight w of m over time t minus hrefA,m SA squared, trying to keep the feet from just hovering awkwardly.

Rosa: Finally, they have joint limit constraints that penalize predicted robot joint angles falling outside the range of qmin and qmax as Llim equals X summed over t x Xj ReLU(ˆqA,t,j minus qmax,j) squared plus ReLU(qmin,j minus ˆqA,t,j) squared. That keeps the joints within their physical boundaries.

Dev: So the main idea is using these constraints to make sure the generative training process produces physically plausible contact patterns and joint movements rather than just mathematically consistent but physically impossible motions.

Taro: But I have to ask, how robust is this when things get messy in real-world scenarios? For example, what happens if the environment changes unexpectedly?

Rosa: That’s a good point, Taro. The paper shows it's validated on the Unitree G1 platform and that the results beat some of the representative baselines in both semantic fidelity and physical realism.

Dev: But we have to think about latency and real-time execution. If this framework is used for teleoperation or imitation learning, what are the practical limits on how fast these generators can run?

Taro: The authors mention that it's designed to reduce dependence on robot-specific paired datasets, but they also note a limitation: it relies on CycleGAN objectives and static target poses as priors to synthesize plausible motions.

Rosa: So, while the framework is unsupervised for motion retargeting across different embodiments, its performance is tied to the quality of those starting cycle objectives and the fixed T-poses they use for normalization.

Dev: That means if you try to retarget a motion onto a robot that has a very unusual or non-standard skeleton, or if the initial cycleGAN training isn't good, you’re going to see artifacts in the final output.

Taro: What about those contact artifacts they are trying to reduce? The paper states they are aiming to mitigate things like foot skating, body floating, and ground penetration by adding those physics-aware constraints.

Rosa: They claim that the explicit physics-aware training objectives make a real contribution to the overall performance, and the experimental results show improvements in physical realism compared to prior methods.

Paper summary: Dev: From my side, I'm focused on how well this translates to control loops. The paper doesn't detail the specific loop rate or latency requirements for deployment, which is something I need clarity on for a safety case.

Taro: That’s a real gap in their description—the paper shows what it does in terms of motion transfer fidelity and physical realism, but it doesn't specify the real-time performance guarantees needed for deploying this on hardware.

Rosa: So, to wrap up the summary of Human2Humanoid: Physics-Aware Cross-Morphology Motion Retargeting for Humanoid Robots, they proposed an unsupervised method using CycleGANs and skeleton-aware GCNs to transfer human motion to humanoid robots without paired data.

Dev: They achieved this by using a morphology-invariant end-effector consistency loss and incorporating explicit physics-aware feasibility constraints—foot contact, height, and joint limits—to improve physical plausibility.

Taro: The implication is that we can use a single human motion capture to generate plausible motions for many different robot bodies, even if the robot shapes are vastly different from the original source motion.

Rosa: It moves us away from needing massive datasets for every single robot model, which is a huge step for practical applications in teleoperation and imitation learning.

Dev: The paper’s title really captures the core idea: it’s not just about retargeting motion; it’s about doing that with physical constraints across different body shapes.

Taro: It suggests that structure-aware learning, like using graph convolutions on skeletons, is a way to handle cross-embodiment problems without needing perfect matches for every single case.

Rosa: The work shows that combining topological feature capture with physics constraints can actually yield better results than methods that only focus on matching the skeleton structure alone.

Dev: So, while it’s not a complete solution for deployment yet, it lays a solid foundation by focusing on preserving motion semantics under large scale mismatches.

Taro: The limitation I see is that the performance still depends heavily on how well those initial cycle objectives are designed, and they don't fully address the real-time guarantees needed for immediate control tasks.

Rosa: That’s a fair caveat. It’s a powerful tool for generating diverse motion ideas, but the next step would be rigorously testing its latency and failure modes in dynamic environments where it needs to operate continuously.

Dev: Exactly. We need to know if this framework can handle unexpected world behavior gracefully without causing control loops to fail or generate unstable joint commands.

Taro: That’s what we’ll be watching for—how it recovers when the physics-aware constraints are challenged by external forces, not just during the training phase.

Conclusion: Rosa: So, we're looking at this paper called Human2Humanoid about transferring human motions to humanoid robots without needing paired data for every single body type.

Dev: Right, so they built this system using a CycleGAN and a graph convolutional network to handle those different body shapes and capture the motion features based on the skeletal structure.

Rosa: And they’ve added these physics-aware constraints—foot contact, foot height, and joint limits—to make sure the resulting motions actually look like something you could run on a real robot.

Dev: That’s what I was focusing on; making sure the system doesn't generate impossible movements like feet skating or joints going through their limits.

Rosa: So, when you look at the title Human2Humanoid: Physics-Aware Cross-Morphology Motion Retargeting for Humanoid Robots, it really sums up what they did.

Dev: It’s about taking human motion and adapting it across different robot shapes while keeping things physically plausible, which is a big deal.

Rosa: The authors are showing that you don't necessarily need tons of specific data for every single robot model if you use this kind of framework to handle the differences between bodies.

Dev: They’re using the cycle gan architecture and those topology-aware features to bridge that gap in a way that respects both the motion structure and the physical reality.

Rosa: It suggests that we can generate motions for many different robots just from one human motion capture, which opens up possibilities for imitation learning a lot more broadly.

Dev: But we still have to figure out how fast this system can run on actual hardware and what happens when the world throws us a curveball outside of the training data.

Rosa: That’s exactly where we need to look next—how robust is this framework when things get messy in real-world scenarios, not just in the controlled lab settings.

More episodes

← Home