A humanoid robot developed by a Japanese robotics company demonstrated advanced dexterity by sorting black socks from a moving conveyor belt carrying mixed black and white socks.

The demonstration took place at South Korean AI developer RLWRLD’s “Dexterity Night” in San Francisco, showcasing advances in robot control powered by its new model, RLDX-1.

Using cameras and human-like hands, the robot identified colors, grasped the socks, and placed them into separate containers while remembering previously detected colors.

Experts say the model showcases South Korea’s growing strength in humanoid robotics, especially in advanced fingertip manipulation technology.

Humanoid dexterity race

A humanoid robot developed by the Japanese robotics company Enactic demonstrated advanced object-handling skills at the Exploratorium event venue in San Francisco by sorting black socks from a conveyor belt carrying a mix of black and white socks, reports The Chosun Daily (TCD).

The robot identified colors using onboard cameras, picked up the socks on a conveyor belt with its hands, and placed the black ones into a separate container. Company representatives said the system could also remember previously identified colors, enabling it to combine judgment, reflexes, and memory during tasks.

The event featured robots from multiple companies, including Japan’s Enactic, South Korea’s WIRobotics, and the US’ Origami Robotics, all operating using RLDX-1. When objects or hands were presented, the robots moved their fingers to grasp items like humans or precisely picked up small objects.

Recently, RLWRLD announced that it is accelerating the deployment of RLDX-1-based systems by collecting training data from multiple real-world sites. The company records skilled human work, including hotel staff, logistics workers, and convenience store employees, using cameras and sensors to build datasets for robot learning. This data is used to enable tasks such as folding, grasping, and organizing in real environments.

According to TCD, beyond RLWRLD, domestic firms are focusing on fingertip technology for humanoid robot commercialization. While the United States leads in high-performance AI models and China in hardware cost competitiveness, Korean companies are focusing on precision driving, force and torque sensing, and manipulation AI.

Another Korean player, Robotis, develops robot hands using a direct-drive method that transmits motor power directly to finger joints without gears or cables and has reportedly received pre-orders from global technology companies, including Google and Apple. Edin Robotics develops sensor technology that replicates fingertip sensation.

Humanlike robot control

RLWRLD’s dexterity-first foundation model, RLDX-1, is designed for high-precision robotic manipulation using high-degree-of-freedom (DoF) robotic hands. The model targets complex real-world industrial tasks that demand fine motor control, continuous contact awareness, and long-horizon decision-making.

According to RLWRLD, current foundation models often fall short in key areas such as contextual memory and force sensing. RLDX-1 addresses these limitations through a full-stack robotics pipeline that integrates scalable data collection, advanced model architecture, and unified training and deployment strategies.

At its core is the Multi-Stream Action Transformer (MSAT), which processes vision, motion, memory, and torque signals in separate streams before merging them to generate coordinated actions. The system also incorporates a robotics-focused vision-language model, physics and motion modules, and a cognition interface that compresses perceptual inputs into memory tokens for long-term task tracking.

To improve learning diversity, RLDX-1 uses a synthetic data engine alongside human hand motion-capture pipelines, expanding its coverage of dexterous manipulation scenarios. RLWRLD reports that the model achieves state-of-the-art performance in both simulated and real-world benchmarks, outperforming leading vision-language-action systems in spatial reasoning, temporal understanding, and contact-rich tasks. It is designed to generalize across multiple robotic embodiments, including single-arm, dual-arm, and humanoid platforms, enabling more adaptive industrial manipulation capabilities.