Robai 与 LinkerBot 合作，推动机器人自监督学习平台 Robot Academy 商业化

가

The body content of the article will be changed to this letter size.

스크롤 이동 상태바

Robai CEO Jung Won-young (left) and LinkerBot co-founder Zhang Yanbai pose after signing a memorandum of understanding. (Photo: Robai)

South Korean physical artificial intelligence (AI) startup Robai is seeking to commercialize a self-supervised learning platform for robots.

Robai announced on Oct. 19 that it has partnered with Chinese robotic hand manufacturer LinkerBot to validate AI control technology that enables robots to learn their own physical structures.

Robai will apply its AI self-exploration platform, Robot Academy, which is currently under development, to LinkerBot’s commercial robotic hardware. The companies plan to assess whether the control technology can be implemented efficiently across different robot architectures.

In physical AI, self-supervised robot learning allows robots to learn from visual, sensor and motion data collected through interactions with real-world environments, without requiring humans to label every action or demonstrate each movement. For example, robots can learn the properties of objects and the conditions required for successful manipulation by analyzing their attempts to grasp objects and the resulting outcomes.

This approach can significantly reduce reliance on conventional robot training data collection methods, including teleoperation and kinesthetic teaching.

Teleoperation involves humans directly controlling robots through remote-control devices, joysticks, motion-capture equipment or virtual reality and haptic interfaces. Data generated during these operations is used for training. Camera footage, joint angles, positions, gripper states and control commands are recorded in synchronization, allowing robots to learn tasks such as object grasping and component assembly from demonstrations.

Kinesthetic teaching involves humans physically guiding robotic arms to generate training data. Although intuitive and well suited to demonstrating precise movements, it places a substantial physical burden on operators.

Both approaches provide real-world robot data but require human operating time and equipment costs. Training performance also depends on the diversity of the collected demonstrations.

To address these limitations, training robots with synthetic data generated in digital twin environments has recently become increasingly mainstream.

Self-supervised learning can further reduce the need for data collection through physical manipulation or synthetic data generation through simulation. However, no commercialized products are available yet.

Google DeepMind has developed AutoRT, a system in which AI plans robot tasks, coordinates their execution and collects training data in real-world environments. Agility Robotics has demonstrated AutoRT in warehouse operations involving the movement of tote boxes, while Figure AI has unveiled technology that converts visual information acquired by robots into physical movements through its proprietary Helix AI model. However, these developments are not yet considered commercialization of self-supervised learning.

“Teleoperation and digital twins are widely used to train robot control for specific tasks, whereas Robai’s AI self-exploration is a method of learning task-independent control and can replace conventional teleoperation and digital twin approaches,” a Robai official said.

Through the partnership, Robai plans to determine whether AI learning can reduce the repetitive control software development and calibration work required whenever robotic hardware changes.

LinkerBot, headquartered in Beijing, manufactures multi-jointed robotic hands. The company previously said it had reached cumulative shipments of 10,000 units in December 2025 and secured more than 80% of the global market for high-degree-of-freedom robotic hands.

Initial technical validation will cover three LinkerBot products: the Linker Hand O6, Linker Hand L20 Lite and Linker Arm P7.

Hyun-Seon, Park hspark@thelec.kr

My comment