ASUS Unveils Gallop Humanoid Robot to Pack Motherboard Accessories on Its Own Factory Line
ASUS factory robot Gallop runs on Jetson AGX Thor, with customer pilots planned for 2027 and sales in 2028

ASUS has unveiled Gallop, its first humanoid robot: a two-armed, fixed-base system that is learning to pack motherboard accessories at a workstation inside an ASUS group electronics factory. Co-CEO S.Y. Hsu introduced the robot in a CNBC interview on November 13, and the company followed with a detailed ASUS Pressroom post by Dr. Albert Chang describing what Gallop does today, how it learns, and what remains unfinished. The announcement marks the PC maker's entry into factory humanoid robotics less than a year after the Gallop program began in early 2026.
An ASUS Factory Line Becomes Gallop's First Training Ground
The choice to put Gallop to work inside an ASUS group plant before offering it to customers is not incidental. It reflects what may be the most strategically important aspect of the program. Real-world manipulation data for electronics manufacturing tasks is difficult to obtain, and a robotics company without its own production lines generally has to arrange access to someone else's factory to gather demonstrations and test on live equipment. ASUS already owns that environment, and it presents the ability to test on its own lines as an advantage.
Gallop's first job is a single station at an ASUS group electronics manufacturing services plant, where a board arrives roughly every 20 seconds and 15 small accessories, many sealed in thin, clear bags, have to go into the box. ASUS says its target is a 99.5% success rate within that 20-second window, which it describes as the same standard it holds a human worker to. That figure is a goal, not a measured result. The company has not published a success rate, independently evaluated benchmark data, or basic specifications such as payload, degrees of freedom, or price.
Gallop comes from the Humanoid Robotics business unit of the ASUS Physical AI Solutions Business Group. Taiwan's TechNews reported in late August that ASUS told employees in an internal memo that its AIoT Business Group would be renamed the Physical AI Solutions Business Group on September 1, combining three divisions: AI solutions, autonomous platforms, and humanoid robots.
A Four-Layer Architecture Designed to Outlast Its AI Models
One of the more technically considered design choices in Gallop is a modular software stack that separates the robot's AI decision-making from its physical control.
ASUS describes four layers. A planner looks at the scene, interprets the job, and breaks it into steps. A supervisor, which ASUS describes as a second AI, checks every step and re-plans when something goes wrong. A skills layer handles reaching, grasping, and placing. A control layer moves the arms and hands precisely and stops the robot immediately if something is off. The layers communicate through fixed interfaces, which ASUS says lets it swap in newer AI models without rebuilding the robot.
This decoupling matters because AI models are improving faster than robot hardware is refreshed. A manufacturer that tightly couples its decision-making model to its motion control code faces a rebuild each time a meaningfully better model arrives. The layered approach treats the AI components as replaceable software rather than as fixed characteristics of the machine.
Everything runs on a single NVIDIA Jetson AGX Thor computer inside the robot. According to NVIDIA, the platform pairs a Blackwell-architecture GPU with a 14-core Arm CPU and 128 GB of LPDDR5X memory at 273 GB/s of memory bandwidth, delivering up to 2,070 FP4 TFLOPS of AI compute in a 40-to-130-watt power envelope. NVIDIA says that amounts to 7.5 times the AI compute and 3.5 times the energy efficiency of the earlier Jetson AGX Orin, a relevant figure for a robot that may eventually work long production shifts. When NVIDIA made the platform generally available in August 2025, it named Agility Robotics, Boston Dynamics, and Figure among the early adopters, and it positions Thor as hardware for its Isaac GR00T humanoid foundation models. ASUS already had experience with the module: its ASUS IoT unit announced the PE3000N edge AI platform built on Jetson Thor in October 2025.
How Gallop Learns to Pack a Motherboard Box
The learning pipeline ASUS describes follows a now-familiar pattern in manipulation robotics: seed the system with human demonstrations, multiply those demonstrations in simulation, refine through direct human guidance, and keep improving with data from the production line.
Experienced workers, with their consent, wear a small head camera so the robot can observe how the job is actually done. Simulation then generates variations of each recorded example. An operator wearing a VR headset and gloves guides the robot through the task using the same parts and cameras as the real station. Once the robot is deployed, what it records feeds back into training.
ASUS says it does not build its own large AI model. It starts from leading open models that already capture how hands and objects behave, then adapts them with its own factory data, an approach meant to reduce how much data each new task requires. The company has not named the models it uses.
The limiting factor is the hand. ASUS designs the robot's body and a dexterous hand with a sense of touch in-house, sourcing motors and other standard components from partners, and it calls the hand the hardest part of any humanoid and one that is still scarce worldwide. That hand is not yet on the robot: ASUS says it will bring it into Gallop over the coming year. Until then, the system's ability to hit its stated target on thin, flexible, bagged parts remains unproven. Hsu acknowledged the difficulty on CNBC, saying the small bagged accessories require special training to handle well, according to Humanoids Daily's account of the interview.
Where Gallop Fits in the Factory Humanoid Race
Factory humanoid programs are developing along different hardware philosophies: mobile systems designed to move through warehouses and plants, and stationary dual-arm workstations designed for precise tasks at a defined position. Gallop is in the second category today, limited to pick-and-place work at one station, though ASUS plans to put it on wheels over the coming year.
The closest available comparison is AGIBOT's G2, a wheeled humanoid that the Chinese company said in April it had integrated into tablet production lines at Longcheer Technology's Nanchang plant. In their joint announcement, the companies reported a success rate above 99.9% and more than 140 hours of cumulative continuous operation, with cycle times of roughly 19 to 20 seconds. Those figures are company-reported, and the task, moving tablets through test stations, is not directly comparable to bagged-accessory packing. Still, the Longcheer deployment is the nearest public benchmark for humanoids on consumer electronics lines.
Foxconn has pursued a self-deployment approach similar in spirit to ASUS's. A Foxconn senior robotics engineer said at NVIDIA's GTC conference in March that the company planned to put "fewer than 10" wheeled humanoids on a server production line in Houston in mid-April to support assembly of NVIDIA GB300 AI servers, according to Build Better. Foxconn has not published results from that pilot. Elsewhere, Agility Robotics signed a commercial agreement in February to deploy seven Digit humanoids at Toyota Motor Manufacturing Canada's Woodstock plant and names Schaeffler among the companies deploying its humanoids. BMW has moved Figure AI's Figure 03 into logistics sequencing work at its Spartanburg plant in South Carolina. Tesla told shareholders in its third-quarter update that Optimus production has begun in Fremont.
Regulation is also reshaping the field. In late July, the Federal Communications Commission moved to ban imports of new foreign-made humanoid and quadruped robots, citing national security risks, in a step widely reported as aimed at China. The restriction limits the options of US manufacturers that might otherwise have considered AGIBOT or similar Chinese systems. Whether it helps ASUS is less clear: as reported, the ban covers foreign-made robots generally, ASUS is headquartered in Taiwan, and the company has not named the partners that will pilot Gallop.
Targets on Paper, Pilots in 2027, Market Availability in 2028
The clearest limitation in the ASUS announcement is that every performance figure is prospective. The 99.5% success rate within 20 seconds is a stated target, not a result reproduced by an independent evaluator or confirmed in a disclosed internal trial. The dexterous hand is designed but not yet integrated. The wheeled version does not yet exist. ASUS says it will show the upgraded robot to customers at Computex 2027, with customer pilots planned for 2027 and market availability targeted for the second half of 2028. Hsu told CNBC that Gallop is scheduled to be released to partners in the second half of 2027.
The market it is entering is growing but crowded. MarketsandMarkets estimates the global humanoid robot market at $5.41 billion in 2026 and projects $50.27 billion by 2035, a 28.1% compound annual growth rate. That forecast reflects industry expectations rather than proven revenue at scale. Counterpoint Research estimates that about 16,000 humanoid robots were installed worldwide in 2025 and that China accounted for more than 80% of those installations, which makes the competitive pressure on newer entrants like ASUS real and immediate.
What ASUS has that many humanoid developers do not is a production environment where Gallop can fail, be corrected, and improve on a task that matters to the electronics manufacturers it hopes to sell to. If the company can integrate its in-house hand and validate it against the speed and success targets before customer pilots begin, it will enter commercial conversations with factory data that is hard for outsiders to replicate. The milestone to watch is not the 2028 availability date but whether results from the 2027 pilots show that the own-factory training loop delivers the performance ASUS is currently targeting.



