Bionic Human-Motion Style Transfer for Physically Executable Whole-Body Control of Humanoid Robots
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "Bionic Human-Motion Style Transfer for Physically Executable Whole-Body Control of Humanoid Robots".
Dev: Here's the gist of the paper. Expressive whole-body motion is important for humanoid robots operating in human environments, where robots are expected to move stably while presenting readable and adjustable body behaviors<ref:2606.03536#pg2> The proposed method achieves a favorable balance between style-transfer quality and robot-oriented physical feasibility<ref:2606.03536#pg2> Problem Formulation The paper studies the problem of generating a full-body reference that preserves the content trajectory, expresses the target human motion style, and can be stably executed by a real humanoid robot<ref:2606.03536#pg2> This formulation introduces several coupled challenges, including how style amplification can alter contact rhythm and center-of-mass motion, which may reduce balance robustness<ref:2606.03536#pg2> Furthermore, diffusion-based generation may produce foot sliding, ground penetration and high-frequency joint jitter if contact and smoothness are only checked after sampling<ref:2606.03536#pg2> The paper aims to connect human-motion-inspired style transfer with robot execution through a physics-aware multi-condition latent diffusion model and a preview-based whole-body tracking policy<ref:2606.03536#pg2> Generation Framework The generation framework follows a unified pipeline from human-motion style generation to humanoid execution<ref:2606.03536#pg2> This pipeline synthesizes a stylized motion using a physics-aware multi-condition latent diffusion model, converts the generated motion into a G1-compatible reference, and executes it on the robot with a whole-body tracking policy<ref:2606.03536#pg2> The generation model is conditioned on content, style and trajectory so that the produced motion can preserve the target action, express the human-derived style and maintain the global path structure<ref:2606.03536#pg2> Training Signal Construction The training signal is built using a physics-aware multi-condition latent diffusion model that fuses content, style and trajectory conditions as explicit condition tokens<ref:2606.03536#pg2> The denoiser organizes content, style and trajectory as explicit condition tokens and fuses them with motion latent tokens through cross-attention<ref:2606.03536#pg2> Classifier-free guidance is enabled by randomly dropping the style condition with probability pdrop during training to adjust the style intensity without retraining<ref:2606.03536#pg2> Contact-consistency and temporal-smoothness regularization are computed on decoded motions during training to reduce artifacts that would hinder downstream execution<ref:2606.03536#pg2> Control and Deployment The control stage involves a reference-preview whole-body tracking policy trained with a cluster-and-distill strategy to handle style-dependent reference variations<ref:2606.03536#pg2> The tracker is trained with both human motion data and generated stylized trajectories, allowing it to adapt to style-dependent reference variations without switching controllers<ref:2606.03536#pg2> The deployment protocol involves an offline stage where the style-transfer model generates a stylized human-motion reference, which is converted and filtered into a G1-compatible whole-body reference<ref:2606.03536#pg2> In the online stage, only the reference-preview tracking policy and the low-level PD controller run on the robot<ref:2606.03536#pg2> Experimental Validation The experiments validate that the proposed method achieves a 96.0% success rate over 125 reported real-robot trials<ref:2606.03536#pg2> The robot-oriented group evaluates physical feasibility using metrics like Foot Sliding Factor (FSF) and measures reference smoothness with Apos and Jpos<ref:2606.03536#pg2> Table 4 Ablation of physics-aware regularization terms shows that imposing both contact regularization and smoothness regularization yields the best values across FSF, Apos and Jpos<ref:2606.03536#pg2> The results show that the proposed method shows a more favorable trade-off, with lower FSF, smoother joint references and a higher simulated closed-loop SR compared to baselines<ref:2606.03536#pg2> Increasing the guidance scale λ strengthens human-like expressive cues but also increases the risk of contact inconsistency and tracking difficulty<ref:2606.03536#pg2> The real-robot study achieves a 96.0% success rate over 125 guidance-scale trials, showing the feasibility of transferring short human style exemplars to real humanoid whole-body motion<ref:2606.03536#pg2> The paper presents a bionic generation-to-control framework for physically executable humanoid motion style transfer<ref:2606.03536#pg2> The proposed method achieves a favorable balance between style-transfer quality and robot-oriented physical feasibility<ref:2606.03536#pg2> Although animation-oriented baselines obtain better kinematic perceptual scores in some reported metrics, the proposed pipeline improves stance-foot stability, joint smoothness and hardware execution success<ref:2606.03536#pg2> The reported Unitree G1 experiments achieve a 96.0% success rate over 125 guidance-scale trials, showing the feasibility of transferring short human style exemplars to real humanoid whole-body motion<ref:2606.03536#pg2> The future work will extend the framework to broader real-robot style-content combinations and richer human–robot interaction scenarios<ref:2606.03536#pg2> Tracking Reward Design The expert policies are optimized with a weighted reference-tracking reward designed to encourage robust tracking of stylized references<ref:2606.03536#pg2> The total reward is the weighted sum of the terms in Table 7, which includes joint-position tracking, joint-velocity tracking, root pose tracking, and end-effector tracking<ref:2606.03536#pg2> The term ρslip penalizes stance-foot slip by measuring horizontal sliding per contact frame<ref:2606.03536#pg2> The term ρlim imposes soft joint-limit regularization to suppress over-extended postures during tracking<ref:2606.03536#pg2> Failure Analysis and Limitations Several limitations are observed, including that excessive style amplification can produce high-frequency jitter and over-extended postures, indicating the guidance scale must be bounded by robot kinematics and tracking robustness<ref:2606.03536#pg2> Strong upper-body swing may shift the center of mass and reduce the support margin when the style exemplar contains exaggerated arm motion<ref:2606.03536#pg2> Asymmetric styles can disturb contact rhythm and cause stance-phase inconsistencies<ref:2606.03536#pg2> A further limitation is possible content-style entanglement, where the MotionCLIP style embedding may contain not only expressive cues but also content-related motion tendencies<ref:2606.03536#pg2> The current real-robot validation is mainly conducted on flat indoor ground and uses a limited number of hardware trials compared to large-scale simulation evaluation<ref:2606.03536#pg2> Style Intensity Analysis The classifier-free guidance scale λ controls the intensity of the transferred style, and as λ increases, the generated motion shows stronger style cues reflected by higher SRA and qualitative body-motion changes<ref:2606.03536#pg2> At the same time, excessive style amplification may increase foot sliding, joint jitter and tracking failure at larger guidance scales<ref:2606.03536#pg2> Figure 4 visualizes the effect of λ in both simulation and real-robot deployment, showing that larger λ values make the transferred style more recognizable but also increase the risk of contact inconsistency and tracking difficulty<ref:2606.03536#pg2> The hardware success rate is 96.0% at guidance scale 0.75, while it drops to 84.0% at guidance scale 1.0<ref:2606.03536#pg2> Tracking Reward Design The expert policies are optimized with a weighted reference-tracking reward where the joint-position tracking term has a weight of 0.30 and the root height, roll and pitch tracking term has a weight of 0.16<ref:2606.03536#pg2> The end-effector tracking for feet and hands is given a weight of 0.16, while stance-foot slip suppression is weighted at 0.06<ref:2606.03536#pg2> The action-rate smoothing term has a weight of 0.03, and sequence completion under fall and safety checks has a weight of –1<ref:2606.03536#pg2> The weights are kept fixed for all expert policies and all tracker-ablation variants<ref:2606.
Title and authors: Rosa: So we're looking at this paper by Huang and colleagues called "Bionic Human-Motion Style Transfer for Physically Executable Whole-Body Control of Humanoid Robots." Rosa, what's the main takeaway for us right now?
Dev: Well, it seems like they're tackling that big problem where we want robots to move stably in human environments but still look expressive, and they’re proposing a way to reuse short human motion styles across different robot movements.
Taro: I wonder if this reuse part is the most practical thing here. We already have ways to transfer style, but usually it's either for simulation or just for animation; making it physically executable is the real hurdle.
Rosa: Exactly, and what they are doing with this paper is proposing a whole generation-to-control pipeline that connects human motion style transfer directly with robot execution. It’s not just about making pretty videos; it’s about getting something the robot can actually do in the real world.
Dev: They use a physics aware multi condition latent diffusion model for generating these stylized motions, and then they feed that into a preview based whole body tracking policy to make sure it executes correctly on the robot.
Taro: That sounds like they are trying to solve that coupling problem, where amplifying the style might mess up how the robot's center of mass moves or how it contacts the ground, which Rosa mentioned earlier.
Rosa: Right, and this is a key area where they focus: they introduce contact consistency and temporal smoothness regularization during training of the diffusion model itself to reduce artifacts like foot sliding or high frequency joint jitter before anything is even executed.
Dev: That sounds like a good way to handle those issues we see when you just check contact after the motion is already sampled, which can lead to foot sliding or ground penetration.
Taro: But the paper also points out that there’s this trade-off: stronger stylization might make it look more human but it can degrade support foot stability or hardware tracking.
Rosa: That trade-off is central to their approach, and they address it by using classifier free guidance, which lets you adjust the style intensity without retraining the generation model by just changing a scale, lambda.
Dev: So you can dial in how much of that human style you want to transfer before running it on hardware? That’s useful for experimentation.
Taro: But they also flag that increasing that guidance scale lambda makes the risk of contact inconsistency and tracking difficulty go up, and they showed a specific point where the hardware success rate drops from ninety-six percent down to eighty-four percent when the scale goes up to one point zero.
Rosa: That's a concrete piece of data showing that there’s a clear limit to how much style you can transfer before it becomes physically unexecutable for this setup.
Title and authors: Dev: And they mention that the preview based whole body tracking policy is trained using a cluster and distill strategy on style diverse motion references, which means the robot doesn't need retraining every time you want a new style intensity or content.
Taro: That addresses that point about having to retrain for every new style; it’s supposed to let the tracker adapt robustly without needing constant retraining for each specific variation.
Rosa: So, overall, this paper presents this generation-to-control pipeline as a way to connect human motion style transfer with robot execution in a way that balances quality and physical feasibility. It aims to use short human style exemplars as reusable bionic sources for physically executable expressive humanoid motion.
Dev: If we look at the results, they achieved a ninety-six percent success rate over one hundred twenty-five reported real-robot trials on the Unitree G1 robot, which shows that this method is feasible outside of just simulation <ref:2606.03536#pg2,success rate over 125 reported real-robot trials>.
Taro: That's a solid number for real-world deployment; it moves beyond just animation benchmarks and shows it works when things get messy.
Rosa: So, to wrap up this paper by Huang et al., they’ve shown that by using a physics aware multi condition latent diffusion model, you can generate motions that preserve the content trajectory while expressing human style, and they've built a system around it for robust execution on hardware.
Dev: The key improvement here seems to be integrating those regularization terms during training to handle the physical constraints early on, which helps reduce those artifacts we worry about in real-time loops.
Taro: I think the most interesting implication is how they formalized that style transfer process—turning it into a reusable source rather than just a one-off generation task.
Rosa: Yeah, and for anyone working on humanoid robotics, this suggests that leveraging short human motion exemplars as a bionic source could be really useful for quickly prototyping and exploring diverse expressive behaviors.
Dev: So while they have those limitations, like the style amplification risk causing jitter at higher guidance scales, the balance they struck between style recognizability and physical executability is quite favorable for practical use cases right now.
Taro: It's a good balance if you’re aiming for things that are physically executable while still looking human-like.
Rosa: That’s what we have here, connecting the generation aspect with the control aspect so it doesn't just look good on paper but actually works on a robot.
Dev: It sets a new direction for how we might approach imitation learning for complex whole body movements in real-world scenarios.
Taro: We should keep watching how they extend this to even broader style and content combinations, because that’s where the real challenge lies.
The paper's summary: Rosa: So we've been looking at how they take those human motion styles—like a specific way someone walks or dances—and try to make that style actually work on a robot, and this paper lays out their whole generation-to-control pipeline for it.
Dev: Yeah, it’s about taking that human movement idea and building a system so the robot can execute it stably while still looking like the human did, which is the main challenge we've been talking about.
Taro: It’s interesting because they tackle this coupling problem head-on where trying to amplify that style—making it look more intense—can actually mess up the robot’s balance or how its center of mass moves.
Rosa: Exactly, and they address that by using a physics aware multi condition latent diffusion model for generating the motion, which means they feed in the content, the style, and the trajectory as explicit tokens to keep everything physically plausible from the start.
Dev: That’s smart because it lets you condition the generation at a token level, not just on some abstract style vector that might ignore gravity or friction rules.
Taro: And they bake in some safety checks during training too, like contact consistency and temporal smoothness regularization to get rid of artifacts like foot sliding before we even test it on hardware.
Rosa: It’s a unified pipeline then: first, you generate the stylized motion using that diffusion model conditioned on all three things—content, style, and trajectory—then you convert it into something the robot can actually use and run with a whole body tracking policy.
Dev: And for execution, they use a preview based whole body tracking policy that’s trained on both human motion data and these generated stylized trajectories, which means the robot learns to adapt to style changes without needing a full retraining every time you want a slightly different intensity.
Taro: That sounds like it handles those variations in style better than just having one static policy for everything.
Rosa: The whole point is that they’re aiming for this bionic generation-to-control framework where you can reuse short human motion styles across different robot contents, making it possible to have a robot perform a variety of expressive tasks based on quick style exemplars.
Dev: It’s about finding that sweet spot between high quality in terms of human look and the physical feasibility needed for real hardware execution.
Taro: But they did flag some limitations too—they noted that if you push the style intensity too high, you can get joint jitter and tracking failure, especially with the guidance scale lambda getting bigger.
Rosa: They showed a concrete point where that guidance scale at one point caused the hardware success rate to drop from ninety-six percent down to eighty-four percent.
Dev: So it’s a trade-off: stronger style cues mean better human look, but you risk making the motion too physically demanding for the robot's joints and stance stability.
Taro: It also mentioned that exaggerated upper body swing in a style exemplar can shift the center of mass and reduce the support margin when you actually try to run it on a real robot.
Rosa: So, they’re showing that while their method is very effective for transferring short human style exemplars to real humanoid whole body motion, there are still limits based on robot kinematics and tracking robustness.
Dev: It really shows that even with these sophisticated AI models, the physical realities of robotics—like contact rhythm changes due to asymmetric styles—are still hurdles that need careful tuning.
Taro: And they also touched on a potential issue where the style embedding might accidentally contain content tendencies, meaning you’re transferring more than just the style you want.
Rosa: So, in short, it’s a solid framework that connects generation to control for physically executable motion transfer, but you still have to be careful with how much style intensity you push.
Dev: It gives us a way to explore human-inspired motion on robots without having to perfectly model every single physical interaction from scratch.
Taro: Next time we talk about how this works, we'll look at those reward designs they used for the tracking policies that actually guide the robot correctly through those tricky execution phases.
The paper's improvements: Tom: So we've been talking about how they generate human motion styles for robots that can actually move in real environments, and now we're going to look at what the authors suggest as improvements to their system.
Rosa: They focus on making sure that when you do this style transfer, you aren't just getting a pretty animation; they’re suggesting adding contact consistency and temporal smoothness regularization during the training of that diffusion model itself.
Dev: That makes sense because those specific checks for foot sliding and joint jitter are what we see in the real world when we run things on hardware, so baking those checks into the training signal is a smart move for improving hardware executability.
Taro: I like that idea of checking contact rhythm before it even runs, because usually you only check those things once you've already sampled the motion and it’s too late.
Rosa: Exactly, and they also suggest using a preview based whole body tracking policy trained with a cluster-and-distill strategy to handle different style variations robustly.
Dev: That’s interesting because it means the robot gets smarter about adapting to style changes without needing us to retrain the entire tracker every single time we want a slightly different intensity or content motion.
Taro: It sounds like they are trying to build a more adaptable execution layer for these stylized references.
Rosa: The implication here is that this isn't just a generation tool; it’s a system where the control policy is designed to be resilient to the stylistic variations that the generator might produce.
Dev: It moves the focus from just making one perfect motion to building a robust controller that can handle a range of plausible, human-like motions coming out of an AI generator.
Taro: And they also hinted at something about how content and style embeddings might get tangled together—meaning the style cues could accidentally carry tendencies related to the specific action content.
Rosa: That’s a real limitation they pointed out, suggesting that future work needs to figure out how to disentangle those two aspects cleanly so the robot doesn't get confused.
Dev: So, it’s about improving the decoupling between what you want the robot to do—the content—and *how* you want it to do it—the style.
Taro: It shifts the challenge from just generating a good look to building a system that understands the physics of both simultaneously.
Rosa: They are laying out a path for future work where they extend this framework to handle even broader combinations of real-world styles and content, which is where the next big test will be.
Dev: It sets up the next round of research to see if this generation-to-control structure holds up when we introduce more complex, messy interactions in a real environment.
Conclusion: Tom: So we're wrapping up our look at this paper, "Bionic Human-Motion Style Transfer for Physically Executable Whole-Body Control of Humanoid Robots." Essentially, they show us a method that connects generating human motion styles to actual robot execution in a way that balances the desired human quality with physical possibility.
Rosa: Right, it’s about taking those complex human movement ideas and building a pipeline so you can reuse short style exemplars across different robot contents while making sure the whole thing is physically sound for real hardware.
Dev: I think what this means for us in controls is that we don't have to start from scratch every time we want a new expressive gait; we just condition the execution policy on a learned style reference.
Taro: It’s cool that they focused on making those regularization terms during training, because it shows they weren't just hoping the motion would look good and then checking it later.
Rosa: They did, and that's what makes it useful for field trials; if the system fails outside the lab, we know exactly which physical constraints—like contact consistency—need strengthening in our training signal.
Dev: And they showed real-world validation with the Unitree G1 robot achieving a ninety-six percent success rate over one hundred twenty-five trials across different guidance scales, which is a solid number for deployment feasibility.
Taro: I'm still thinking about that warning about the guidance scale lambda increasing and then the hardware success rate actually dropping when you go past one point; it tells us there’s a very clear physical ceiling to how much style we can inject without breaking things.
Rosa: That limit is important because it tells us we have to respect those robot kinematics and tracking robustness boundaries when pushing for more expressive styles.
Dev: So, the implication is that this approach gives us a way to explore human-inspired motion without completely ignoring the latency and failure modes of real-time loops, provided we stay within those physical bounds.
Taro: It opens up avenues for creating robot behaviors that feel more natural and expressive, rather than just stiff or robotic motions.
Rosa: Yeah, this paper really sets a foundation for using short human motion styles as reusable bionic sources for more varied robot interactions in the future.
Dev: Before we move on, we should keep an eye on how they address those limitations regarding content-style entanglement and see if that gets solved in later work.
Taro: Definitely, because untangling those two things is going to be a big part of making these systems truly flexible for complex real-world scenarios.
Tianchen Huang, Mingkuan Zhao, Yang Gao, Feiyang Yuan, Junchi Gu, Xiaohu Zhang, Dongdong Zhao, Shi Yan
Institute of Humanoid Robots, Department of Precision Machinery and Precision Instrumentation, University of Science and Technology of China · School of Computer Science and Technology, Faculty of Electronic and Information Engineering, Xi’an Jiaotong University · School of Information Science and Engineering, Lanzhou University
cs.RO
Submitted: 2026-06-02
Updated: 2026-06-02
Comments: Project page: https://huangtc233.github.io/bionic-style-transfer/
Journal ref: Journal of Bionic Engineering 23(6), 3708-3723 (2026)
DOI: 10.1007/s42235-026-00986-8
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 90/100
The gist: Expressive whole-body motion is important for humanoid robots operating in human environments, where robots are expected to move stably while presenting readable and adjustable body…
Key concepts
- Physics-aware multi-condition latent diffusion model
- This is the core generation engine that creates the stylized motion. It takes three inputs—the desired content (what the person is doing), the style (how they are moving), and the trajectory (the path)—and fuses them using cross-attention within a diffusion process. This allows it to generate motions that look human while respecting basic physical constraints.
- Reference-preview whole-body tracking policy
- This is the control system that makes the robot move correctly. It is trained using a 'cluster-and-distill' approach on many different styles. This allows the robot to adapt smoothly to variations in style without needing to be retrained every time a new style or content is requested.
- Classifier-free guidance scale (λ)
- This parameter controls how strongly the desired human style is applied during motion generation. Increasing this scale makes the robot's movement look more like the target human style, but too high can cause instability, like foot sliding or joint jitter on the robot.
- Contact-consistency and temporal-smoothness regularization
- These are safety checks performed during training to ensure generated motions are executable. They specifically penalize artifacts that cause problems when the motion is played back on a robot, such as feet slipping or joints vibrating too quickly, leading to smoother, more stable results.
Terminology
Summary
Expressive whole-body motion is important for humanoid robots operating in human environments, where robots are expected to move stably while presenting readable and adjustable body behaviors<ref:2606. p. 0 cn;
†These authors contributed equally to this work. Abstract
Expressive whole-body motion is important for humanoid robots operating in
human environments, where robots are expected to move stably while presenting readable and adjustable body behaviors. The generated references are then converted
into G1-compatible robot references and executed by a preview-based whole-body
tracking policy trained with a cluster-and-distill strategy. Simulation and Unitree G1 experiments show that the proposed method can transfer short human
style exemplars to diverse robot motion contents, reduce contact and jitter artifacts compared with animation-oriented style-transfer baselines, and achieve a
96. The generated references are then converted
into G1-compatible robot references and executed by a preview-based whole-body
tracking policy trained with a cluster-and-distill strategy. Simulation and Unitree G1 experiments show that the proposed method can transfer short human
style exemplars to diverse robot motion contents, reduce contact and jitter artifacts compared with animation-oriented style-transfer baselines, and achieve a
96.Where the paper says it
Where the paper says it
Where the paper says it
Problem Formulation
The paper studies the problem of generating a full-body reference that preserves the content trajectory, expresses the target human motion style, and can be stably executed by a real humanoid robot<ref:2606.03536#pg2> This formulation introduces several coupled challenges, including how style amplification can alter contact rhythm and center-of-mass motion, which may reduce balance robustness<ref:2606. p. 2 This formulation introduces several coupled challenges. First, style amplification
can alter contact rhythm and center-of-mass motion, which may reduce balance robustness. First, style amplification
can alter contact rhythm and center-of-mass motion, which may reduce balance robustness. Second, diffusion-based generation may produce foot sliding, ground penetration
and high-frequency joint jitter if contact and smoothness are only checked after sampling. The method is not intended as a pure animation generator or
a pure controller trainer. Instead, it connects human-motion-inspired style transfer
with robot execution through a physics-aware multi-condition latent diffusion model,
retargeting and a preview-based whole-body tracking policy.Where the paper says it
Where the paper says it
Where the paper says it
Generation Framework
The generation framework follows a unified pipeline from human-motion style generation to humanoid execution<ref:2606.03536#pg2> This pipeline synthesizes a stylized motion using a physics-aware multi-condition latent diffusion model, converts the generated motion into a G1-compatible reference, and executes it on the robot with a whole-body tracking policy<ref:2606. p. 4 The goal
is to transfer the demonstrated style to the given content motion, instead of replaying
the style exemplar as a fixed scripted behavior. The pipeline synthesizes a stylized motion using a physics-aware multi-condition
latent diffusion model, converts the generated motion into a G1-compatible reference, and executes it on the robot with a whole-body tracking policy. The pipeline synthesizes a stylized motion using a physics-aware multi-condition
latent diffusion model, converts the generated motion into a G1-compatible reference, and executes it on the robot with a whole-body tracking policy. The generation
model is conditioned on content, style and trajectory so that the produced motion
can preserve the target action, express the human-derived style and maintain the
global path structure.Where the paper says it
Where the paper says it
Training Signal Construction
The training signal is built using a physics-aware multi-condition latent diffusion model that fuses content, style and trajectory conditions as explicit condition tokens<ref:2606.03536#pg2> The denoiser organizes content, style and trajectory as explicit condition tokens and fuses them with motion latent tokens through cross-attention<ref:2606. p. 5 The difference lies in how these conditions are fused and optimized
for humanoid execution. Instead of relying only on content concatenation and adaptive
modulation of self-attention and feed-forward layers, the proposed denoiser organizes
content, style and trajectory as explicit condition tokens and fuses them with motion
latent tokens through cross-attention. The denoiser ϵθ is
trained with a mean squared error objective as
Ldiff = Ex0,τ,ϵ h
∥ϵ − ϵθ(xτ , τ, C, csty, ctraj)∥
2
2
i
. (3)
Classifier-free guidance is enabled by randomly dropping the style condition with probability pdrop during training. The generation
model is conditioned on content, style and trajectory so that the produced motion
can preserve the target action, express the human-derived style and maintain the
global path structure. Contact-consistency and temporal-smoothness regularization
are applied during generation training to reduce artifacts that would hinder downstream execution.Where the paper says it
Where the paper says it
Where the paper says it
Control and Deployment
The control stage involves a reference-preview whole-body tracking policy trained with a cluster-and-distill strategy to handle style-dependent reference variations<ref:2606.03536#pg2> The tracker is trained with both human motion data and generated stylized trajectories, allowing it to adapt to style-dependent reference variations without switching controllers<ref:2606. p. 4 Contact-consistency and temporal-smoothness regularization
are applied during generation training to reduce artifacts that would hinder downstream execution. The tracker is trained with both human motion data and generated
stylized trajectories, allowing it to adapt to style-dependent reference variations without switching controllers. 3 Deployment Protocol
The proposed framework is deployed in two stages. In the offline stage, the styletransfer model generates a stylized human-motion reference, which is converted and
10 This
design allows the policy to anticipate style-dependent rhythm and contact timing,
which is important for stabilizing the robot near swing–stance transitions. The policy
outputs target joint positions at 50 Hz, which are tracked by a low-level PD controller running at 1000 Hz.Where the paper says it
Where the paper says it
Where the paper says it
Experimental Validation
The experiments validate that the proposed method achieves a 96.0% success rate over 125 reported real-robot trials<ref:2606. p. 1 Simulation and Unitree G1 experiments show that the proposed method can transfer short human
style exemplars to diverse robot motion contents, reduce contact and jitter artifacts compared with animation-oriented style-transfer baselines, and achieve a
96. 0% success rate over 125 reported real-robot trials. Instead, it connects human-motion-inspired style transfer
with robot execution through a physics-aware multi-condition latent diffusion model,
retargeting and a preview-based whole-body tracking policy. Contact-consistency and
temporal-smoothness regularization are computed on decoded motions during diffusion training, so that style cues are encouraged to remain within a robot-executable
motion space before tracking. The generated references are then converted
into G1-compatible robot references and executed by a preview-based whole-body
tracking policy trained with a cluster-and-distill strategy. Simulation and Unitree G1 experiments show that the proposed method can transfer short human
style exemplars to diverse robot motion contents, reduce contact and jitter artifacts compared with animation-oriented style-transfer baselines, and achieve a
96. We develop a physics-aware multi-condition latent diffusion model that injects style
cues while suppressing contact inconsistency and high-frequency joint oscillations
through decoded-motion regularization. We integrate the generated stylized references with a whole-body tracking policy
trained on style-diverse motion references through a cluster-and-distill strategy,
enabling robust deployment of diverse stylized motions. Rather than only asking whether
a policy can track an existing reference, we study how to generate a short-exemplarbased stylized reference that remains suitable for tracking. The closed loop from
human style exemplar to physics-aware generation, retargeting, tracking and real-robot
validation is the central distinction. 0% success rate over 125 reported real-robot trials. The results demonstrate
the feasibility of using short human motion exemplars as reusable bionic sources
for physically executable expressive humanoid motion. 0% success rate over 125 reported real-robot trials. The results demonstrate
the feasibility of using short human motion exemplars as reusable bionic sources
for physically executable expressive humanoid motion.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
The paper presents a bionic generation-to-control framework for physically executable humanoid motion style transfer<ref:2606.03536#pg2> The proposed method achieves a favorable balance between style-transfer quality and robot-oriented physical feasibility<ref:2606.03536#pg2> Although animation-oriented baselines obtain better kinematic perceptual scores in some reported metrics, the proposed pipeline improves stance-foot stability, joint smoothness and hardware execution success<ref:2606. p. 3 These animation-oriented methods are important baselines, but they rarely guarantee robot executability. A generated motion may look natural or carry recognizable
style cues in kinematic evaluation, while still exhibiting stance-foot sliding, ground
penetration or high-frequency joint jitter that violates hardware constraints. Simulation and Unitree G1 experiments show that the proposed method can transfer short human
style exemplars to diverse robot motion contents, reduce contact and jitter artifacts compared with animation-oriented style-transfer baselines, and achieve a
96. 0% success rate over 125 reported real-robot trials. Given a short human style exemplar and a target content
motion, the proposed framework generates a stylized whole-body reference that
preserves the intended motion content while transferring the demonstrated style. A physics-aware multi-condition latent diffusion model is developed to fuse
style, content and trajectory conditions, and classifier-free guidance is used to
1
arXiv:2606.Where the paper says it
Where the paper says it
Where the paper says it
Tracking Reward Design
The expert policies are optimized with a weighted reference-tracking reward designed to encourage robust tracking of stylized references<ref:2606.03536#pg2> The total reward is the weighted sum of the terms in Table 7, which includes joint-position tracking, joint-velocity tracking, root pose tracking, and end-effector tracking<ref:2606. p. 9 The expert policies are trained with a DeepMimic-style reference-tracking reward,
following the design principle used in recent humanoid motion-tracking systems:
reference-tracking terms are kept positive and dominant, while a small number of regularization terms are added for sim-to-real robustness and smooth control [29, 35, 38]. The reward encourages the simulated robot to track the converted reference in joint
space, root state and key end-effectors, while penalizing stance-foot slip, joint-limit
violation and abrupt action changes. The
reference foot speed is defined by finite differences of joint positions as v
ref
f,t = ∥q
ref
f,t −
q
ref
f,t−1
∥2, and the stance mask is constructed as sf,t = 1[v
ref
f,t < δv] with δv = 0. The
contact loss applies hinge penalties to the predicted foot speed v
pred
f,t = ∥q
pred
f,t −q
pred
f,t−1
∥2
and height h
pred
f,t with a penetration tolerance δh = 0. The expert policies are trained with a DeepMimic-style reference-tracking reward,
following the design principle used in recent humanoid motion-tracking systems:
reference-tracking terms are kept positive and dominant, while a small number of regularization terms are added for sim-to-real robustness and smooth control [29, 35, 38]. The reward encourages the simulated robot to track the converted reference in joint
space, root state and key end-effectors, while penalizing stance-foot slip, joint-limit
violation and abrupt action changes.Where the paper says it
Where the paper says it
Where the paper says it
Failure Analysis and Limitations
Several limitations are observed, including that excessive style amplification can produce high-frequency jitter and over-extended postures, indicating the guidance scale must be bounded by robot kinematics and tracking robustness<ref:2606.03536#pg2> Strong upper-body swing may shift the center of mass and reduce the support margin when the style exemplar contains exaggerated arm motion<ref:2606. p. 3 A stylized gait modifies support timing, trunk posture
and arm swing simultaneously. Therefore, expressive locomotion cannot be treated
as an upper-body animation overlay; it must be designed together with contact
consistency, balance and tracking robustness. The
reference foot speed is defined by finite differences of joint positions as v
ref
f,t = ∥q
ref
f,t −
q
ref
f,t−1
∥2, and the stance mask is constructed as sf,t = 1[v
ref
f,t < δv] with δv = 0. The
contact loss applies hinge penalties to the predicted foot speed v
pred
f,t = ∥q
pred
f,t −q
pred
f,t−1
∥2
and height h
pred
f,t with a penetration tolerance δh = 0. (3)
Classifier-free guidance is enabled by randomly dropping the style condition with probability pdrop during training. Specifically, the MotionCLIP style embedding before the
style projection network is replaced by a zero vector for unconditional noise prediction,
while the content and trajectory conditions are kept unchanged.Where the paper says it
Where the paper says it
Where the paper says it
Style Intensity Analysis
The classifier-free guidance scale λ controls the intensity of the transferred style, and as λ increases, the generated motion shows stronger style cues reflected by higher SRA and qualitative body-motion changes<ref:2606.03536#pg2> At the same time, excessive style amplification may increase foot sliding, joint jitter and tracking failure at larger guidance scales<ref:2606. p. 2 First, style amplification
can alter contact rhythm and center-of-mass motion, which may reduce balance robustness. Second, diffusion-based generation may produce foot sliding, ground penetration
and high-frequency joint jitter if contact and smoothness are only checked after sampling. Simulation and Unitree G1 experiments show that the proposed method can transfer short human
style exemplars to diverse robot motion contents, reduce contact and jitter artifacts compared with animation-oriented style-transfer baselines, and achieve a
96. 0% success rate over 125 reported real-robot trials. 1 Pipeline Overview
A unified pipeline from human-motion style generation to humanoid execution is presented. For any executable content motion from MoCap, video-derived skeletons or
open motion repositories, the user supplies a short human style exemplar together
with a guidance scale λ to control style intensity at inference. cn;
†These authors contributed equally to this work. Abstract
Expressive whole-body motion is important for humanoid robots operating in
human environments, where robots are expected to move stably while presenting readable and adjustable body behaviors. 1 Pipeline Overview
A unified pipeline from human-motion style generation to humanoid execution is presented. For any executable content motion from MoCap, video-derived skeletons or
open motion repositories, the user supplies a short human style exemplar together
with a guidance scale λ to control style intensity at inference. Given a short human style exemplar and a target content
motion, the proposed framework generates a stylized whole-body reference that
preserves the intended motion content while transferring the demonstrated style. A physics-aware multi-condition latent diffusion model is developed to fuse
style, content and trajectory conditions, and classifier-free guidance is used to
1
arXiv:2606.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The expert policies are optimized with a weighted reference-tracking reward where the joint-position tracking term has a weight of 0.30 and the root height, roll and pitch tracking term has a weight of 0. p. 9 The future reference covers a one-second preview window with
10 uniformly sampled future instants. For each preview instant, root height, roll and
pitch, root linear velocity, yaw angular velocity in the local root frame, and joint
references are provided. Given a short human style exemplar and a target content
motion, the proposed framework generates a stylized whole-body reference that
preserves the intended motion content while transferring the demonstrated style. A physics-aware multi-condition latent diffusion model is developed to fuse
style, content and trajectory conditions, and classifier-free guidance is used to
1
arXiv:2606. The expert policies are trained with a DeepMimic-style reference-tracking reward,
following the design principle used in recent humanoid motion-tracking systems:
reference-tracking terms are kept positive and dominant, while a small number of regularization terms are added for sim-to-real robustness and smooth control [29, 35, 38]. The reward encourages the simulated robot to track the converted reference in joint
space, root state and key end-effectors, while penalizing stance-foot slip, joint-limit
violation and abrupt action changes. The expert policies are trained with a DeepMimic-style reference-tracking reward,
following the design principle used in recent humanoid motion-tracking systems:
reference-tracking terms are kept positive and dominant, while a small number of regularization terms are added for sim-to-real robustness and smooth control [29, 35, 38]. The reward encourages the simulated robot to track the converted reference in joint
space, root state and key end-effectors, while penalizing stance-foot slip, joint-limit
violation and abrupt action changes. Given a short human style exemplar and a target content
motion, the proposed framework generates a stylized whole-body reference that
preserves the intended motion content while transferring the demonstrated style. A physics-aware multi-condition latent diffusion model is developed to fuse
style, content and trajectory conditions, and classifier-free guidance is used to
1
arXiv:2606. Simulation and Unitree G1 experiments show that the proposed method can transfer short human
style exemplars to diverse robot motion contents, reduce contact and jitter artifacts compared with animation-oriented style-transfer baselines, and achieve a
96. 0% success rate over 125 reported real-robot trials. During expert training,
mild domain randomization is applied to link mass, link inertia, ground friction, joint
friction, motor response and sensor noise. The same randomization setting is used
for all experts and tracker-ablation variants.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Bionic generation-to-control framework for exemplar-driven style transfer allows AI systems to generate a stylized whole-body reference that preserves content trajectory while transferring human motion style cues, enabling reuse of short human style exemplars across diverse robot motion contents.
-
Physics aware multi condition latent diffusion model integrates content, style and trajectory conditions using explicit condition tokens fused via cross attention, allowing the denoiser to retrieve these cues at the token level for generating physically plausible motions.
-
The system incorporates contact consistency and temporal smoothness regularization during training of the diffusion model to reduce artifacts like foot sliding and high-frequency joint oscillations before execution, ensuring better hardware executability.
-
A preview based whole body tracking policy, trained with a cluster-and-distill strategy on style diverse motion references, enables the robot to adapt robustly to style dependent reference variations without requiring retraining for each new style intensity or content motion.
-
Classifier free guidance allows adjustment of style intensity without retraining the generation model by combining unconditional and conditional noise predictions using a guidance scale lambda, controlling the strength of transferred style at inference time.
Abstract
Expressive whole-body motion is important for humanoid robots operating in human environments, where robots are expected to move stably while presenting readable and adjustable body behaviors. However, most expressive motions are still obtained from fixed demonstrations or manually designed scripts, making it difficult to reuse a demonstrated style across different motion contents. Inspired by the way human motion styles convey affective and intentional cues through gait rhythm, posture, arm swing and body sway, this paper proposes a bionic generation-to-control framework for exemplar-driven style transfer on humanoid robots. Given a short human style exemplar and a target content motion, the proposed framework generates a stylized whole-body reference that preserves the intended motion content while transferring the demonstrated style. A physics-aware multi-condition latent diffusion model is developed to fuse style, content and trajectory conditions, and classifier-free guidance is used to adjust the style intensity without retraining. To improve hardware executability, contact-consistency and temporal-smoothness regularization are imposed on decoded motions during training. The generated references are then converted into G1-compatible robot references and executed by a preview-based whole-body tracking policy trained with a cluster-and-distill strategy. Simulation and Unitree G1 experiments show that the proposed method can transfer short human style exemplars to diverse robot motion contents, reduce contact and jitter artifacts compared with animation-oriented style-transfer baselines, and achieve a 96.0% success rate over 125 reported real-robot trials. The results demonstrate the feasibility of using short human motion exemplars as reusable bionic sources for physically executable expressive humanoid motion.
Sources
- SMooDi: Stylized Motion Diffusion Model
- Diffusion-based Human Motion Style Transfer with Semantic Guidance
- Human Motion Diffusion Model
- MotionDiffuse: Text-Driven Human Motion Generation with Diffusion Model
- Perpetual Humanoid Control for Real-time Simulated Avatars
- Universal Humanoid Motion Representations for Physics-Based Control
- Expressive Whole-Body Control for Humanoid Robots
- GMT: General Motion Tracking for Humanoid Whole-Body Control
- Track Any Motions under Any Disturbances
- ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills
- BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion
- KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills
- KungfuBot2: Learning Versatile Motion Skills for Humanoid Whole-Body Control
- MotionCLIP: Exposing Human Motion Generation to CLIP Space
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation