As the race to mass-produce Humanoid Robot intensifies, the ability to secure data is emerging as a key competitive edge. That is because the volume and quality of data used to train artificial intelligence (AI) are considered the decisive factors that determine a humanoid's performance. Corporations are particularly moving to strengthen their data competitiveness by securing "synthetic data." Synthetic data refers to data artificially generated by AI models, simulations, or statistical methods, rather than data collected from real-world environments.
On the 27th, according to the robotics industry, domestic physical artificial intelligence (AI) company RLWRLD unveiled its self-developed Robotics foundation model, "RLDX-1."
A Robotics foundation model is a general-purpose AI model that learns from large-scale data and can be applied to diverse environments and tasks. Just as large language models (LLMs) learn from massive volumes of text data to generate answers or reason, a Robotics foundation model is designed to learn from real-world data to perform a wide range of physical tasks.
RLDX-1, developed by RLWRLD, is a model designed to allow a robot's hand to move as precisely as a human hand. It can understand not only the weight, shape, and movement of objects but also complex physical environments.
Data is the key factor that determines the performance of a robot foundation model. In particular, for the hands of humanoids deployed on industrial sites to move in line with factory environments, it is important to secure data needed for learning about factory settings, including not only the movements of real human hands but also the positions of objects.
RLWRLD has built a system to independently secure the data needed to train its vision-language model (VLM) and vision-language-action model (VLA) for robots.
In general, to train a robot hand model, robotic devices such as grippers are worn on a person's hand to collect motion data. However, this method can restrict natural human movement due to the equipment and has limitations in being applied to five-finger robot hands similar to human hands.
RLWRLD's RLDX-1 collects data by filming human hands and tracking bare-hand movements. It converts the filmed joint movements of the hand into digital data and uses it for VLA training. Through this method, RLWRLD enabled the robot foundation model to secure the quality and volume of data needed to learn complex, real hand movements, such as adjusting hand pressure according to an object's weight or moving fingers according to specific task goals.
It also established a system to secure data for learning factory environments where robot hands are deployed by using synthetic data.
Data collected from real environments is high in quality but difficult to secure in large quantities and incurs high expense. In contrast, synthetic data can be generated at scale, and by using AI and simulations, it can create data from diverse environments that are hard to obtain in the real world, supplementing the limits of real data.
RLDX-1 has a synthetic data acquisition system that uses video generation AI. An RLWRLD official said, "Robot work environments are often complex and unstructured, making it very difficult to collect real robot data inside factories," adding, "By building a synthetic data system based on a video generation model, we generated new videos that can occur in real situations, including various objects, lighting, and backgrounds, and amplified the size of the dataset by about fivefold."
ROBOTIS recently released a video of its self-developed humanoid, "AI Sapiens," dancing to choreography by a domestic idol group. After learning from choreography videos, AI Sapiens additionally trained on various motion data using synthetic data. Because complex dance moves require precise control not only of joint movements but also of balance and posture transitions, large-scale training data is essential.
A ROBOTIS official said, "To implement complex choreography, we need sufficient training data as well as actuator performance," adding, "Because we secured a large amount of data through synthetic data, we can stably implement highly difficult movements."
Academia is also focusing on using synthetic data to improve humanoid motion performance.
Research teams from four universities, including Fudan University and Shanghai Jiao Tong University, reported in findings posted in early this month on the preprint site "arXiv" that they proposed methods for securing synthetic data using three-dimensional (3D) simulations and training humanoid data.
The researchers explained, "With a synthetic data technique using simulation, we can generate large-scale data that matches real motions," and, "Using 3D models, we reconstructed realistic objects from real images and augmented the data through teleoperation in simulated environments."
They added, "After conducting extensive experiments on actual Humanoid Robot, the results trained with simulation data showed a higher success rate in most tasks than the approach trained with real robot data."
Han Jae-kwon, a professor in the Department of Robotics Engineering at Hanyang University, said, "Just as test scores vary depending on who studies and how, a robot's performance varies depending on the data acquisition system and the quality of the data," adding, "There are dozens of ways to acquire human motion data, and various ideas are emerging from corporations."
He added, "If you collect data through synthetic data, you can generate data that is dangerous or difficult to make, and you can actually create a vast amount of data, on the order of several hundred thousand," explaining, "It won't be easy to build such a data acquisition system, but ultimately that will become the company's capability."