X-Diffusion: Training Diffusion Policies on Cross-Embodiment Human Demonstrations
X-Diffusion employs diffusion models to learn cross-embodiment robot policies from human demonstrations, achieving 16% success rate improvement.
Key Findings
Methodology
This paper introduces X-Diffusion, a framework based on Ambient Diffusion, which treats human demonstrations as noisy versions of robot actions. It unifies state and action spaces, adds Gaussian noise to actions, and trains a classifier to identify the source embodiment. During training, human actions are only incorporated when the classifier is fooled—i.e., at high noise levels—ensuring the policy learns from data that resembles feasible robot actions. The process involves action space alignment, noise injection, classifier training, and selective data fusion, optimizing the diffusion reverse process for cross-embodiment transfer.
Key Results
- X-Diffusion outperformed naive co-training and manual filtering across five manipulation tasks, with an average success rate increase of 16%. For example, in the Pan On Plate task, success improved from 65% to 81%.
- The classifier effectively filtered infeasible human actions, enabling the model to learn from diverse demonstrations without degrading performance.
- The method demonstrated positive transfer, especially in complex tasks, by leveraging high-noise stages of human demonstrations, validating the noise-based approach for cross-embodiment learning.
Significance
This work addresses the critical challenge of utilizing large-scale human videos for robot training, reducing reliance on costly robot data collection. By employing noise diffusion to bridge embodiment gaps, it opens new avenues for scalable robot learning, especially in real-world applications where collecting robot demonstrations is impractical. Its ability to filter and incorporate diverse human behaviors enhances the robustness and generalization of learned policies, impacting both academia and industry in autonomous manipulation and service robotics.
Technical Contribution
The key technical innovation lies in integrating Ambient Diffusion with a dynamic classifier to control the inclusion of human data based on noise levels. This approach allows the model to learn from low-quality, cross-embodiment data without sacrificing kinematic feasibility. It introduces the concept of a minimum indistinguishability step, guiding the selective fusion of human demonstrations. The framework combines diffusion-based denoising with behavior cloning, providing theoretical guarantees on the quality of transferred behaviors and practical improvements over existing methods.
Novelty
This is the first application of Ambient Diffusion to robot cross-embodiment learning, leveraging the gradual noise process to blur embodiment differences. The dynamic classifier-based filtering during diffusion training is a novel mechanism that adaptively determines when human demonstrations can be safely used, surpassing prior approaches that rely on manual filtering or naive data mixing.
Limitations
- The effectiveness depends on accurate action space alignment and classifier performance; errors here can lead to suboptimal data filtering.
- In scenarios with extreme embodiment differences or highly complex tasks, the noise threshold may not fully eliminate infeasible actions, risking degraded policy quality.
- Training involves multiple stages of noise injection and classifier optimization, which increases computational cost and complexity, posing challenges for real-time deployment.
Future Work
Future directions include developing adaptive noise thresholds, integrating multi-modal sensory data for richer action understanding, extending to multi-agent settings, and combining with reinforcement learning for policy refinement. Enhancing scalability and reducing computational overhead are also key goals to facilitate real-world deployment.
AI Executive Summary
Robotics research has long grappled with the challenge of acquiring sufficient high-quality training data. While collecting robot demonstrations is expensive and time-consuming, human videos offer a scalable alternative. However, the fundamental difference in embodiment makes direct imitation infeasible, often leading to the learning of unrealistic or unsafe behaviors. To address this, the authors propose X-Diffusion, a novel framework that leverages diffusion models to bridge the embodiment gap.
X-Diffusion builds upon Ambient Diffusion, a generative modeling approach that introduces Gaussian noise into data, gradually transforming it into a distribution where differences in embodiment are blurred. By treating human actions as noisy counterparts of robot actions, the method employs a classifier trained to distinguish between the two sources. During training, human demonstrations are only incorporated when the classifier is fooled—i.e., at high noise levels—ensuring that only actions resembling feasible robot behaviors influence policy learning.
This approach effectively filters out infeasible behaviors, allowing the policy to learn from a broad spectrum of human data without degradation. Extensive experiments across five manipulation tasks demonstrate that X-Diffusion achieves a 16% average success rate improvement over naive co-training and manual filtering. The results highlight the method’s ability to transfer useful task cues from humans to robots, even when their physical capabilities differ significantly.
The significance of this work lies in its potential to democratize robot training, reducing reliance on costly data collection while improving policy robustness. It opens pathways for scalable, real-world robot learning, especially in complex manipulation scenarios. Future work aims to refine noise thresholds, incorporate multi-modal data, and extend to multi-robot systems, promising further advances in autonomous manipulation technology.
Deep Analysis
Background
机器人自主学习经历了从模仿学习到深度强化学习的演变。早期方法如行为克隆（Behavior Cloning）和逆强化学习（Inverse Reinforcement Learning）在模拟环境中取得一定成功，但在真实世界中面临数据稀缺和成本高昂的问题。近年来，深度学习结合大规模演示数据显著提升了机器人自主能力，但采集高质量机器人示范仍昂贵。人类视频演示因其易于获取，成为潜在的替代方案，但因动作表现差异，难以直接迁移。已有研究尝试通过动作映射、关键点匹配等技术缓解体型差异，但仍存在动作不可行和动态不匹配的问题。生成模型中的Ambient Diffusion提供了处理低质量数据的思路，为跨体型迁移提供了新途径。本文结合这些技术，旨在突破跨体型模仿的瓶颈，充分利用人类演示数据。
Core Problem
核心问题在于如何在跨体型演示中提取有用的任务信息，同时避免学习到不可行的动作。人类演示虽丰富，但因体型和动作方式不同，导致动作不可直接执行，影响策略的有效性。传统联合训练方法在数据融合时容易引入动态和运动学上的不一致，导致策略性能下降。如何筛选和利用这些数据，成为提升机器人自主学习效率的关键难题。解决方案需在保留任务导向的同时，过滤掉不可行的动作，确保学习的动作具有实际可行性。
Innovation
本研究的创新点主要包括：1）将Ambient Diffusion引入机器人跨体型学习，将人类演示作为噪声数据逐步模糊差异；2）训练动作来源分类器，动态控制人类数据的融合阈值，避免不可行动作干扰；3）提出最小不可区分阈值的概念，确保只在噪声足够大时引入不可行动作，从而实现有效迁移。这些创新结合了生成模型和动作识别技术，突破了传统方法的局限，显著提升跨体型迁移的效果。
Methodology
- �� 统一状态与动作空间：利用Grounded SAM 2进行目标对象分割，结合HaMeR检测手部关键点，映射到机器人末端执行器。• 动作噪声引入：在动作序列上加入高斯噪声，逐步生成噪声版本，模拟不同体型差异。• 分类器训练：用带噪声的动作和状态信息训练二分类器，识别动作来源（人类或机器人），以动态调节数据融合阈值。• 最小不可区分阈值：定义在扩散步骤中，分类器无法区分动作来源的最早点，用于筛选人类演示的融合。• 逆扩散训练：在逆过程引入筛选的噪声动作，优化策略，使其在不同噪声水平下学习任务。• 目标优化：结合行为克隆和噪声去噪，确保策略在实际操作中的可行性和鲁棒性。
Experiments
采用五个真实操控任务，分别为关抽屉、放锅、推盘、挂杯和立瓶。每个任务采集5个机器人演示和100个人类演示，利用不同基线（纯机器人、联合训练、筛选后人类数据）进行比较。指标为成功率，进行10次实地测试，统计平均成功率。模型超参数包括扩散步骤数K、噪声系数β，以及分类器训练的平衡参数。还设计了消融实验，验证筛选机制和噪声阈值的影响。
Results
X-Diffusion在五个任务中均优于基线，平均成功率提升16%。在Pan On Plate任务中，从65%提升至81%。筛选机制有效过滤不可行动作，避免策略学到不合理行为。高噪声阶段的人类演示实现正迁移，特别在复杂操作中表现优异。分类器准确识别动作来源，噪声阈值动态调节，确保融合的动作具有实际可行性。整体结果显示，该方法在跨体型迁移中具有显著优势。
Applications
该技术适用于工业机器人、服务机器人等场景，尤其在缺乏大规模机器人演示数据时，利用人类视频快速扩展训练集。只需基本的目标检测和手部关键点识别，即可实现跨体型迁移。未来可结合强化学习，提升策略的适应性和鲁棒性，推动自主操作在复杂环境中的应用。
Limitations & Outlook
模型依赖于动作空间的良好统一和分类器的准确性，若映射不充分或分类失误，可能影响效果。噪声阈值设定在极端差异场景下可能不足以过滤所有不可行动作。训练过程复杂，计算成本较高，实际部署需优化效率。未来需解决多模态融合和自适应阈值问题，以增强泛化能力。
Plain Language Accessible to non-experts
想象你在厨房做饭，手里拿着锅铲，准备炒菜。机器人也是在厨房里，但它的身体结构和你不同，比如没有灵巧的手指，只能用夹子。你用不同的方式炒菜，比如用手翻炒或推锅，机器人也有自己的动作方式。有时候，你会用一种特别的技巧，机器人可能做不到。现在，如果我们让机器人学会你的做菜方法，问题在于它能不能模仿你的动作？
这就像在教一个笨手笨脚的朋友做菜，你告诉他一些技巧，但他不能完全照搬。我们用一种叫“噪声”的魔法，把你的动作变得模糊，让机器人只看到大致的动作轮廓。这样，他就不会学到那些不适合自己身体的细节，而只学到做菜的核心步骤。通过不断调整“模糊”程度，我们让机器人逐渐理解哪些动作是可以模仿的，哪些是不行的。最终，机器人可以学会用自己的方式完成任务，而不用担心模仿得太死板或不切实际。这就像让机器人变得更聪明、更灵活，能在厨房里自如操作。
ELI14 Explained like you're 14
想象你在学校里学做菜，你的老师教你用锅炒菜。你们的身体结构不同，你用手翻炒，机器人用夹子夹东西。你会用不同的动作，但都能做出好菜。现在，假如我们想让机器人学会你的做菜方法，就像你教朋友一样，但问题是，机器人不能完全模仿你的动作，因为它的身体不同。于是，我们用一种特别的技巧，把你的动作变得模糊一些，就像用“模糊魔法”一样，让机器人只看到大概的动作轮廓。这样，机器人就不会学到那些它做不到的细节，只会学到做菜的核心步骤。随着“魔法”变得越来越模糊，机器人逐渐学会用自己的方式完成任务。这样，它就能学会在厨房里灵活操作，不用担心模仿得太死板。这就像让机器人变得更聪明、更会做菜一样！
Abstract
Human videos are a scalable source of training data for robot learning. However, humans and robots significantly differ in embodiment, making many human actions infeasible for direct execution on a robot. Still, these demonstrations convey rich object-interaction cues and task intent. Our goal is to learn from this coarse guidance without transferring embodiment-specific, infeasible execution strategies. Recent advances in generative modeling tackle a related problem of learning from low-quality data. In particular, Ambient Diffusion is a recent method for diffusion modeling that incorporates low-quality data only at high-noise timesteps of the forward diffusion process. Our key insight is to view human actions as noisy counterparts of robot actions. As noise increases along the forward diffusion process, embodiment-specific differences fade away while task-relevant guidance is preserved. Based on these observations, we present X-Diffusion, a cross-embodiment learning framework based on Ambient Diffusion that selectively trains diffusion policies on noised human actions. This enables effective use of easy-to-collect human videos without sacrificing robot feasibility. Across five real-world manipulation tasks, we show that X-Diffusion improves average success rates by 16% over naive co-training and manual data filtering. The project website is available at https://portal-cornell.github.io/X-Diffusion/.