人形机器人借助人类运动数据训练的 AI 学会冲刺与回旋踢

September 10, 2026 feature Humanoid robot learns to sprint and perform spin kicks using AI trained on human motion data by Ingrid Fadelli, Phys.org Ingrid Fadelli Author Meet our staff & contributors Learn about our editorial standards edited by Lisa Lock, reviewed by Robert Egan Lisa Lock Scientific Editor Meet our editorial team Behind our editorial process Robert Egan Senior Editor Meet our editorial team Behind our editorial process Editors' notes This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: fact-checked peer-reviewed publication trusted source proofread The GIST Add as preferred source Overview of the proposed versatile humanoid control framework. (A) Scalable and robust learning from human motions with agile, humanlike behaviors via motion tracking. (B) Versatile control over unseen downstream tasks with diverse learned motor skills via guided diffusion. Credit: Science Robotics (2026). DOI: 10.1126/scirobotics.adx8924 Humanoid robots, robotic systems with body shapes and limbs resembling those of humans, could potentially assist people with manual tasks in various real-world settings. So far, however, most of these robots can reliably perform only a limited set of movements. Teaching humanoid robots to perform a wide range of agile, humanlike motions can be both challenging and time-consuming. One way to do this entails training controllers (i.e., algorithms that determine how a robot moves) using a large pool of human demonstrations. While this approach can allow robots to reproduce some motions, it sometimes yields unnatural movements or requires adjustments tailored to individual actions. In addition, many previously proposed methods worked by training or tuning control policies for specific motions or goals, limiting their ability to generalize to new tasks. Researchers at the University of California, Berkeley (UC Berkeley), and Stanford University recently developed BeyondMimic, a new artificial intelligence framework that could broaden the movement repertoire of humanoid robots while reducing the need for motion-specific tuning or task-specific retraining. The framework, introduced in a paper published in Science Robotics, combines a training approach called reinforcement learning with a model that generates movement plans by refining noise within a compressed representation of motion. The humanoid robot learned a large set of natural and agile humanlike motions and composed them to complete versatile tasks, guided by objectives that were not seen during training. Credit: Science Robotics (2026). DOI: 10.1126/scirobotics.adx8924 How the BeyondMimic framework works To train their AI framework, the researchers used approximately 2.5 hours of motion data showing humans walking, running, dancing, practicing martial arts, jumping, doing cartwheels and performing other agile movements. The movements included in this data were subsequently adapted to the proportions and joints of the G1 humanoid robot developed by Unitree Robotics. Using reinforcement learning, the team trained control policies to reproduce the positions, orientations and speeds of body parts during each reference movement. Reinforcement learning is a widely used method to train computer algorithms through rewards and penalties. In this case, the model was rewarded for accurate motion tracking, while penalties discouraged jerky movements, unsafe joint positions and unwanted contact between the robot's body parts. "We present BeyondMimic, a framework that scales to diverse motions and carries the versatility to compose them seamlessly in tackling unseen downstream tasks," wrote Qiayuan Liao, Takara E. Truong and their colleagues in their paper. "A compact motion tracking formulation enables mastery of a wide range of highly agile behaviors, including aerial cartwheels, spin kicks, flip kicks, and sprinting, with a single setup and shared hyperparameters, all while achieving humanlike performance." To compress the actions learned by the robot, the team used a variational autoencoder, an algorithm that translates action data into a compact and structured representation (i.e., a lower-dimensional latent space). A decoder subsequently converts these compressed representations into commands for the robot. The researchers then trained an algorithm known as a diffusion model to learn sequences containing both the robot's predicted physical states and compressed movement representations. During training, they added noise to these sequences, and the model learned to reconstruct the original movements. "Moving beyond the mere imitation of existing motions, we propose a unified latent diffusion model that empowers versatile goal specification, seamless task switching, and dynamic composition of these agile behaviors," wrote Liao, Truong and their colleagues. "Leveraging classifier guidance, a diffusion-specific technique for test-time optimization toward unseen objectives, our model extended its capability to solve downstream tasks never encountered during training, including motion inpainting, joystick teleoperation, and obstacle avoidance, and transferred these skills zero-shot to real hardware." Credit: Science Robotics (2026). DOI: 10.1126/scirobotics.adx8924 Initial demonstrations and future possibilities The researchers first validated the motions in simulation and then deployed 30 representative motion clips on a Unitree G1 robot. Their findings were promising, as the robot successfully performed various movements, including balancing poses, dancing, crawling, sprinting, kicks, spinning jumps and aerial cartwheels. During these experiments, the researchers also asked 77 participants to compare the walking and running motions produced using BeyondMimic with those generated by Unitree's standard controller. In 70.8% of the comparisons, participants classified BeyondMimic's movements as more humanlike and natural than those enabled by the standard controller. In the future, the team's framework could be refined and tested using an even broader range of motions. It could also potentially be adapted to other humanoid robotic platforms. Written for you by our author Ingrid Fadelli, edited by Lisa Lock, and fact-checked and reviewed by Robert Egan—this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive. If this reporting matters to you, please consider a donation (especially monthly). You'll get an ad-free account as a thank-you. Publication details Qiayuan Liao et al, BeyondMimic: From motion tracking to versatile humanoid control via guided diffusion, Science Robotics (2026). DOI: 10.1126/scirobotics.adx8924 Journal information: Science Robotics Key concepts Embodied robotic manipulationBioinspired soft roboticsAutonomous robotic locomotionHumanoid robotics Who's behind this story? Ingrid Fadelli Freelance journalist with BSc Psychology and MA International Journalism. Covers AI, robotics, neuroscience, and astrophysics since 2018. Full profile → Lisa Lock BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021. Full profile → Robert Egan Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile → © 2026 Science X Network Citation: Humanoid robot learns to sprint and perform spin kicks using AI trained on human motion data (2026, September 10) retrieved 10 September 2026 from https://techxplore.com/news/2026-09-humanoid-robot-sprint-ai-human.html This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only. From cartwheels to backflips, motion-imitation framework teaches three robots dynamic movements Aug 27, 2026 A new approach to reproduce human and animal movements in robots May 5, 2022 Humanoid robots master parkour and acquire human-like agility Mar 5, 2026 Learning dance moves could help humanoid robots work better with humans Jul 11, 2024 New framework enables animal-like agile movements in four-legged robots Jul 13, 2024 Animal-inspired AI robot learns to navigate unfamiliar terrain Jul 11, 2025 AI controller translates VR, video and language commands into humanoid robot actions 21 hours ago A stolen-phone safeguard could disrupt home alarms and block new flagship phones 14 hours ago New Apple CEO unveils latest iPhone lineup, including a foldable model called Duo 16 hours ago New software tool lets users repair AI-generated 3D models, then fabricate them just the way they want 16 hours ago A 60-year-old computing strategy gets an update for AI-era data centers 17 hours ago Ultrablack wool offers a sustainable route to low-cost water purification 17 hours ago