Now recruiting capture teams worldwide. Get paid to record first-person video of real work.

Apply to capture

Glossary

What Is UH-1 / Humanoid-X (Learning from Massive Human Videos for Humanoid Pose Control)?

UH-1 is a language-conditioned humanoid pose model trained on Humanoid-X, a dataset mined from human videos. Firsthand is not affiliated.

Last reviewed: November 2026

Short answer

UH-1 (Universal Humanoid-1) is a Transformer model that turns text commands into humanoid robot pose control. It is trained on Humanoid-X, a dataset of 163,800 motion samples that the authors mined from human videos, captioned, and retargeted to a humanoid. The method is from arXiv 2412.14172; Firsthand is not affiliated.

01

What is UH-1 / Humanoid-X, and what problem does it address?

UH-1 and Humanoid-X are introduced in “Learning from Massive Human Videos for Universal Humanoid Pose Control” (arXiv 2412.14172) by Jiageng Mao, Siheng Zhao, Siqi Song, Tianheng Shi, Junjie Ye, Mingtong Zhang, Haoran Geng, Jitendra Malik, Vitor Guizilini and Yue Wang, with affiliations listed as the University of Southern California, UC Berkeley and Toyota Research Institute. The project page, titled “Universal Humanoid Robot Pose Learning from Internet Human Videos”, lists an oral presentation at the 2025 International Conference on Humanoid Robots. The authors say reinforcement learning and teleoperation for whole-body control are limited by the diversity of simulated environments and the high cost of demonstration collection, while human videos are ubiquitous. Humanoid-X is their attempt to turn such videos into text-labelled humanoid actions, and UH-1 is the model trained on it.

02

Where does Firsthand fit next to UH-1 / Humanoid-X?

UH-1 and Humanoid-X are third-party research. Firsthand is not affiliated with the authors, and no named model, UH-1 included, was trained on Firsthand data. The site states custom egocentric collection to a written spec, fast and on demand, delivering first-person human video with 3D hand pose at 21 joints per hand and a per-joint visibility flag, calibrated stereo depth and 6-DoF IMUs at 200 Hz. The only language data tied to Firsthand capture is one task instruction per episode. Firsthand does not build or train models, and it does not deliver the paper’s captions, parametric body-model fits, humanoid keypoints, robot actions or control policy.

03

How does Humanoid-X turn human videos into humanoid actions?

Per the paper, a fully automated pipeline has five modules. It mines single-person clips from noisy videos, captions the action in each clip, estimates parametric 3D human poses, retargets the motion to a humanoid, and trains a goal-conditioned reinforcement learning policy with PPO to produce physically deployable actions. Each motion sample therefore holds five modalities: the video clip, a text description, human poses, humanoid keypoints and humanoid actions.

The authors mine from three kinds of source: academic datasets for digital-human research, datasets for video action understanding, and YouTube videos found with over 400 search terms, taking the top 20 videos per term. Clips are standardised to 20 frames per second, must show exactly one visible person, and must run at least 64 consecutive frames. Retargeting uses 12 joints shared by humans and humanoids: hips, knees, ankles, shoulders, elbows and wrists.

04

How does UH-1 turn text into humanoid pose control?

UH-1 first discretises humanoid actions into action tokens, which the authors describe as a vocabulary of motion primitives. A Transformer then conditions on a text embedding and auto-regressively decodes those tokens. A proportional-derivative controller converts the resulting joint targets into motor torques.

The paper describes two interchangeable control modes. Text-to-keypoint generates high-level humanoid keypoints for the goal-conditioned policy to track in closed loop. Text-to-action generates robot actions for direct open-loop control; on the real robot the authors found this works only for the upper body, so a pretrained locomotion policy controls the lower body.

05

What did the authors report, and what are the limits?

On the HumanoidML3D benchmark the authors report an over 23% reduction in FID against two text-to-motion baselines, and that pre-training on Humanoid-X improves FID from 0.445 to 0.379 against training on HumanoidML3D alone. They also report a success rate of nearly 100% across 12 language commands, each evaluated 10 times, on a real humanoid robot, and averages above 89% for both control modes in simulation. These are the authors’ figures on their own setups; Firsthand has not verified them.

The authors state that the paper studies humanoid pose control only. Humanoid manipulation is out of scope, and they list learning humanoid loco-manipulation from Internet videos as future work.

06

What hardware, licence and ethics terms does the paper state?

The paper names one robot for real-world testing, the Unitree H1-2. It names no capture rig, because the human data is existing video. Firsthand does not claim to use that robot.

The ethics statement says the Internet videos are strictly for academic research and not intended for commercial use, that face anonymisation was applied to human subjects in the Internet videos, and that the original Internet videos are not released. The licence for the code, dataset and weights is Not stated in the sources read, so check the repository and dataset pages before any commercial use.

07

How does UH-1 differ from HumanPlus, OKAMI, HAT and motion retargeting?

UH-1 sits beside several entries on this site. HumanPlus shadows a person’s motion from an RGB camera, OKAMI imitates a manipulation skill from one RGB-D video, and HAT co-trains on an egocentric human-humanoid dataset. Motion retargeting is a step inside the Humanoid-X pipeline rather than a competing method. The first table below sets them side by side.

08

What does Firsthand state it can supply for this kind of work?

UH-1 learns from Internet video that the authors mined and processed. Firsthand’s stated scope is different: custom collection of first-person human video, depth, 3D hand pose and one task instruction per episode, confirmed first on a Spec Pilot. Robot trajectories, retargeted actions and action labels are not Firsthand deliverables, and the site makes no claim that its data suits any UH-1 pipeline.

Where coverage is deep, the median from signed spec to first validated batch is 48 hours; where it is thin, three to four weeks. Billing is per validated hour. Formats are RLDS, LeRobot format or WebDataset under a buyer-owned commercial license.

09

UH-1 and Humanoid-X at a glance, as stated in the paper

UH-1 and Humanoid-X at a glance, as stated in the paper
ItemWhat the paper states
PaperLearning from Massive Human Videos for Universal Humanoid Pose Control (arXiv 2412.14172)
Method nameUH-1, Universal Humanoid-1, a large humanoid model for language-conditioned pose control
DatasetHumanoid-X: 163,800 motion samples, over 20 million frames, approximately 240 hours
Modalities per sampleVideo clip, text description, parametric 3D human poses, humanoid keypoints, humanoid actions
Video sourcesAcademic digital-human datasets, video action-understanding datasets and YouTube (over 400 search terms, top 20 videos each)
ModelTransformer over discrete action tokens, text-conditioned, auto-regressive; two modes: text-to-keypoint and text-to-action
Real robot namedUnitree H1-2
Stated scope limitHumanoid pose control only; manipulation not studied
Code and dataOfficial GitHub repository and Hugging Face collection; licence Not stated in the sources read
Firsthand relationshipNone; no named model was trained on Firsthand data

11

UH-1 beside related entries on this site

UH-1 beside related entries on this site
Site entryWhat the entry statesHow UH-1 / Humanoid-X differs (sources read)
Humanoid foundation modelA generalist robot policy for a human-shaped body, pretrained on web, human, synthetic and robot data, then adapted to a specific humanoidUH-1 is a text-conditioned pose-control model trained on video-derived humanoid actions; the paper does not call it a foundation model
HumanPlusA system where a simulation-trained policy lets a humanoid shadow a person’s body and hand motion from an RGB cameraUH-1 takes text commands instead of a live person and learns from mined Internet video
OKAMIGenerates a manipulation plan from a single RGB-D human video using object-aware retargetingUH-1 covers pose control from text, not manipulation from one demonstration video
Humanoid Policy ~ Human Policy (HAT)Pairs the PH2D egocentric human-humanoid dataset with a policy whose state-action space is unified across humans and humanoidsHumanoid-X is mined from existing videos, with humanoid actions produced by retargeting and reinforcement learning
Motion retargetingMaps captured human motion onto a robot with different kinematicsRetargeting is one module in the Humanoid-X pipeline, applied to 12 shared joints

12

What the Humanoid-X pipeline produces beside what the Firsthand site states

What the Humanoid-X pipeline produces beside what the Firsthand site states
AspectHumanoid-X (sources read)What the Firsthand site states
Source footageMined single-person clips from academic datasets and YouTubeCustom first-person human video, collected to a written spec
LanguageA text description generated per clipOne task instruction per episode
Body signalParametric 3D human poses estimated from video, then humanoid keypoints3D hand pose at 21 joints per hand with a per-joint visibility flag
Depth and inertialNot listed as inputsCalibrated stereo depth and 6-DoF IMUs at 200 Hz
Robot actionsHumanoid actions produced by a reinforcement learning policyNot delivered; Firsthand captures first-person human video, not robot or teleoperation data
Training useThe authors’ own training and evaluationNo named model was trained on Firsthand data; no affiliation
DeliveryPublic code and dataset pages; licence Not stated in the sources readSpec Pilot, then billed per validated hour; custom collection, fast and on demand

13

Frequently asked questions

What is UH-1?
UH-1, Universal Humanoid-1, is a published Transformer model (arXiv 2412.14172) that takes text commands and outputs humanoid robot actions for pose control. It is third-party research.
What is the Humanoid-X dataset?
Per the paper, Humanoid-X holds 163,800 motion samples, over 20 million frames and approximately 240 hours, mined from human videos. Each sample has a video clip, a text description, human poses, humanoid keypoints and humanoid actions.
How does UH-1 do humanoid pose control from human videos?
The authors caption mined clips, estimate 3D human poses, retarget them to a humanoid and train a reinforcement learning policy for deployable actions. UH-1 then learns to decode action tokens from text.
Did the authors test UH-1 on a real robot?
Yes. The paper reports tests on a Unitree H1-2 with 12 language commands, each evaluated 10 times, and a success rate of nearly 100%. These are the authors’ figures; Firsthand has not verified them.
Is UH-1 or Humanoid-X open source, and can it be used commercially?
The authors link an official GitHub repository and a Hugging Face collection. The licence is Not stated in the sources read, and the paper says the Internet videos are for academic research, not commercial use. Read the licence text first.
Was UH-1 trained on Firsthand data?
No. Firsthand is not affiliated with the authors and claims no use, training or compatibility. Captions, body-model fits, humanoid keypoints, robot actions and control policies are not Firsthand deliverables.
Does UH-1 handle manipulation?
No. The authors say the paper studies humanoid pose control only and that humanoid manipulation is out of scope, listing loco-manipulation from Internet videos as future work.
How can I request a custom egocentric dataset for humanoid learning?
Request a custom dataset with your objects, interactions and signals. Collection is fast and on demand, with a first validated batch in a median of 48 hours where coverage is deep, confirmed first on a Spec Pilot.