Shanghai, Aug. 31, 2026 (GLOBE NEWSWIRE) — Maniformer, a leading Physical AI data platform, today announced two major milestones in real-world data infrastructure for Physical AI: the accumulation of more than one million hours of real-world human data and the delivery of its 20,000th MEgo-series data collection device.
Together, the milestones highlight the growing demand for scalable infrastructure that can collect, process, evaluate, and deliver high-quality real-world data for Physical AI.
Surpassing 1 Million Hours of Real-World Data
As Physical AI systems move beyond controlled environments and into homes, workplaces, factories, retail stores, restaurants, warehouses, and other real-world settings, access to diverse, high-quality data is becoming one of the industry’s biggest bottlenecks.
Unlike traditional robot teleoperation, Maniformer’s MEgo series enables human behavior and manipulation activities to be captured directly from real-world environments, without requiring a robot to be present during data collection.
Maniformer has now accumulated more than one million hours of real-world data, spanning:
- 22 major scene categories
- 10,000+ real-world environments
- 50,000+ object categories
- 500+ fine-grained tasks
- 8 major application areas, including residential, office, retail, manufacturing, warehousing, food service, entertainment, and transportation
The dataset combines synchronized RGB, Depth, IMU, tactile, audio, and motion trajectory data, providing multimodal representations of human actions, objects, and environments.
Maniformer collects real-world data through three complementary configurations:
- Ego View — first-person data capture for bare-hand interactions using a head-mounted device
- Ego + Wrist — synchronized wrist-mounted sensing for hand poses and continuous motion trajectories
- UMI Gripper — end-effector data for robot manipulation and gripper trajectories
Of the more than one million hours collected, approximately 500,000 hours come from Ego View, 350,000 hours from Ego + Wrist, and 150,000 hours from UMI Gripper.
By capturing natural human behavior across diverse environments, these systems create a broader source of real-world training data for Physical AI models without relying exclusively on robot teleoperation.
From Real-World Data to Spatial Understanding
Collecting data in the real world is only the first step. Human actions are difficult to reconstruct accurately, particularly in in-the-wild environments characterized by fast movements, hand motion, severe occlusion, and diverse interactions
To address this challenge, Maniformer has developed MEgo Engine, its data processing infrastructure covering data processing, perception, annotation, and quality evaluation.
At the event, the company introduced HandPose, its high-precision hand reconstruction technology. The system combines five-camera perception, robust hand tracking, multi-view depth fusion, SOTA 3D hand reconstruction, and trajectory recovery to reconstruct hand motion in challenging real-world conditions.
According to Maniformer’s internal evaluations, HandPose achieved state-of-the-art results across seven core metrics, with PA-MPJPE reconstruction error reduced by 62% compared with the next-best solution and frame-to-frame jitter reduced by 93%.
Maniformer’s unified MPR (MEgo Physical Reconstruction) algorithm stack reconstructs the spatial trajectories of the head, hands, and robot grippers with a trajectory error of less than 1 cm, according to the company’s internal evaluations.
These capabilities are designed to turn unstructured real-world interactions into structured spatial and motion data that can be used for model training and evaluation.
Building a Data-to-Model Feedback Loop
For physical AI, the value of data ultimately depends on whether it can improve model performance.
Maniformer has developed ManiEval, a multidimensional data quality evaluation and grading system that assesses data across dimensions including sensor quality, semantics, task characteristics, and motion quality.
Instead of simply filtering data into “good” and “bad” samples, ManiEval uses a quality-based grading approach, allowing data to be matched to specific model training and evaluation requirements.
The company is also developing technologies that connect real-world data with model training and deployment. Its REMORA task-progress model helps identify a robot’s progress through a task and detect execution errors, while its Value Model is used to identify potentially valuable failure data from real-world deployments.
This creates a continuous loop:
Collect → Reconstruct → Evaluate → Train → Deploy → Learn
Data collected from the real world can improve models; model deployment generates new feedback and failure cases; and high-value feedback can then be incorporated into the data pipeline for further training.
Maniformer is currently providing its real-world data services to leading AI and robotics organizations, supporting a range of development needs across embodied AI and vision-language-action (VLA) models.
Scaling the Data Infrastructure for Physical AI
The physical AI industry is moving toward increasingly capable models that must understand and act in the physical world. Scaling Physical AI requires not only more capable models and robots, but also a scalable supply of diverse, high-quality real-world data.
With 20,000 MEgo devices delivered and more than one million hours of real-world data accumulated, Maniformer aims to build an infrastructure layer connecting real-world experiences with the next generation of physical AI models.
“One million hours and 20,000 devices are important milestones for Maniformer, but we see them as the beginning of something much larger,”said Maoqing.“Physical AI will ultimately require tens of billions of hours of real-world experience. Our goal is to transform that experience into high-value data at scale and help accelerate the arrival of the AGI era.”
—
About Maniformer
Maniformer is a leading Physical AI data platform building the data infrastructure for embodied intelligence. Combining intelligent data-collection hardware, a powerful data engine, and end-to-end data services, Maniformer transforms real-world physical interactions into high-quality, scalable data. From collection and processing to validation and intelligent operations, Maniformer provides the data foundation that accelerates the development and deployment of Physical AI.
Find out more about Maniformer here: https://maniformer.ai/en
CONTACT: Linko Song media@maniformer.ai
