Holiday Robotics
Last reviewed
May 9, 2026
Sources
15 citations
Review status
Source-backed
Revision
v5 · 2,934 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Last reviewed
May 9, 2026
Sources
15 citations
Review status
Source-backed
Revision
v5 · 2,934 words
Add missing citations, update stale details, or suggest a clearer explanation.
| Holiday Robotics | |
|---|---|
| General information | |
| Korean name | 홀리데이로보틱스 |
| Founded | April 2024 |
| Founder & CEO | Song Ki-young (송기영) |
| Headquarters | Gangnam-gu, Seoul, South Korea |
| Industry | Robotics, Artificial intelligence |
| Products | Friday wheeled humanoid robot |
| Total funding | 17.5 billion KRW (~$12.9 million, seed round) |
| Lead investor | Stonebridge Ventures |
| Website | holiday-robotics.com |
Holiday Robotics (Korean: 홀리데이로보틱스) is a South Korean robotics company headquartered in the Gangnam district of Seoul. It was founded in April 2024 by Song Ki-young (송기영), the serial entrepreneur best known for co-founding SuaLab, a deep learning-based vision inspection company that was acquired by American industrial vision leader Cognex in 2019 for approximately $195 million in what was at the time the largest foreign acquisition of a Korean technology startup. Holiday Robotics develops humanoid robots for industrial automation, with its flagship product being Friday, a wheeled humanoid that prioritizes dexterous manipulation over bipedal locomotion. The company's stated mission is "liberating people from physical labor with versatile robots."[1][2][3]
The company emerged from stealth in August 2024 with a 17.5 billion KRW (approximately $12.9 million) seed round led by Stonebridge Ventures, the largest seed round raised by any South Korean startup in the first eight months of 2024. Holiday Robotics unveiled the first public prototype of Friday on October 30, 2025, eighteen months after its founding, and announced plans to begin small-batch industrial production in 2026 at a target list price of roughly 100 million KRW (about $75,000) per unit.[4][5][6]
Song Ki-young co-founded SuaLab in 2013 in Seoul as a developer of deep learning-based machine vision software for factory defect detection and quality inspection. Operating in the same broad domain as conventional rule-based machine vision but using convolutional neural networks rather than hand-engineered features, SuaLab built a customer base across Korean and East Asian manufacturing before being acquired by Cognex on October 18, 2019. Cognex paid approximately $195 million in cash for the company, a transaction widely reported in the Korean press as the largest M&A exit ever achieved by a Korean technology startup at that time.[7][8]
At the time of the acquisition Song held a 25.54 percent stake in SuaLab. Stonebridge Ventures, which had backed the company since 2015 with roughly 13 billion KRW across multiple rounds, owned about 17 percent and reported an internal rate of return of approximately 105.79 percent on its position. After the deal closed Song joined Cognex as the head of its deep learning engineering division, where he served a four-year contractual employment period before leaving to start his next venture. He has cited the experience of running a deep-learning-driven product organization inside a global industrial vision company as a key influence on Holiday Robotics' decision to focus on physical manipulation rather than perception alone.[7][9]
Holiday Robotics was incorporated in April 2024, immediately after Song completed his employment obligations at Cognex. He recruited a founding team of AI software engineers and hardware controller specialists drawn from the Korean robotics ecosystem and from former colleagues at SuaLab and Cognex. In an early interview Song framed the new company as a continuation of his earlier work, saying that at SuaLab he "saw the potential of industrial AI" and now wanted to "challenge implementing that intelligence in physical space, the moving intelligence."[2][9]
On August 27, 2024, just over four months after incorporation, Holiday Robotics announced that it had closed a seed round of 17.5 billion KRW (approximately $12.9 to $13.2 million depending on the exchange rate cited). Stonebridge Ventures led the round, with participation from SpringCamp, Atinum Investment, InterVest, and Hyundai Motor's open-innovation vehicle Zer01ne. The round set a record as the largest seed financing for a South Korean startup over the first eight months of 2024 and one of the largest disclosed seed rounds in Korean robotics history. Choi Dong-youl, chief investment officer at Stonebridge Ventures, justified the bet by noting that the founding team possessed "talent in both AI software and hardware controller development, which is critical for humanoid robots."[4][5]
At the time of the announcement Holiday Robotics stated that it would deploy the proceeds primarily to hire specialists in robot control, reinforcement learning, large language models, and simulation engineering, and to establish an overseas robotics research center in the United States during the first half of 2025 to recruit international talent.[5][6]
For its first eighteen months Holiday Robotics operated largely in stealth, restricting public communications to occasional founder interviews and partnership announcements. The company signed a strategic memorandum of understanding on August 1, 2025 with Endotlight (엔닷라이트), a Korean generative 3D AI startup led by Park Jin-young, to jointly produce 3D synthetic data for robot training. Under the agreement Endotlight supplies datasets generated by its Trinix synthetic data platform and proprietary 3D modeling engine, and Holiday Robotics applies them to perception and motion training for industrial scenarios. The two companies described the partnership as an effort to reduce the time and cost required to construct training data for physical AI systems.[10]
On October 30, 2025 the company unveiled the first working Friday prototype to media. The launch presentation emphasised that every core subsystem (the body structure, the actuators, the dexterous hands, and the magnetic tactile sensors) had been designed in-house from a clean sheet rather than assembled from off-the-shelf components, an approach Song described as building the robot "from the ground up." Holiday Robotics stated that it intended to demonstrate Friday performing industrial tasks before the end of 2025 and to begin pilot deployments with manufacturing customers in 2026.[1][11]
In September 2025 Song delivered a public lecture in which he described what he called the four open challenges that humanoid robots must overcome before reaching mass commercialization: slow learning (citing reports that Tesla Optimus had required engineers to repeat motions up to 20,000 times for data collection), the gap between simulation success and real-world reliability, safety risks of operating near humans (he noted that industrial robots cause an average of about 2.7 worker fatalities per year in Korea), and price. He singled out price as "the biggest stumbling block," arguing that while Chinese competitor Unitree advertises a $16,000 base unit, fully optioned configurations can add tens of thousands of dollars in tactile sensors and other accessories, with single-finger force sensors alone reportedly priced at around $14,000. He concluded that humanoids will not see mass adoption until their landed cost approaches that of an automobile.[12]
Friday is Holiday Robotics' first product and the platform on which the company plans to enter commercial service. Unlike many humanoid platforms that emphasise bipedal walking, Friday is a wheeled mobile manipulator: it uses a differential-drive wheeled base instead of legs, a design choice the company argues delivers better stability, longer battery life, lower energy consumption, and predictable safety behaviour, all at substantially lower cost than a comparable biped. The product name reflects an internal roadmap in which weekday names denote distinct platforms: Friday for industrial work, with later "Saturday" and "Sunday" variants planned for service and household applications respectively.[1][13]
| Specification | Value |
|---|---|
| Height | 176 cm |
| Weight | 115 kg total (including 66 kg mobile base) |
| Total degrees of freedom | 63 |
| Arm DOF | 7 per arm |
| Torso DOF | 5 |
| Hand DOF | 20 per hand |
| Hand weight | ~500 g per hand |
| Per-finger-joint load | up to 5 kg |
| Per-arm payload | 5 kg |
| Maximum proximal payload | 20 kg |
| Maximum base speed | 1.9 m/s (~6.84 km/h) |
| Battery | Hot-swappable, 24/7 operation capable |
| Locomotion | Differential drive wheeled base |
| Onboard compute | NVIDIA Jetson Orin |
| Operating system | Linux |
| Connectivity | Wi-Fi, cloud connectivity |
Friday's body has 63 degrees of freedom in total, distributed across seven-DOF arms, a five-DOF torso, and a head and gripper assembly. The total mass is 115 kg, of which 66 kg is concentrated in the mobile base; this front-heavy configuration gives the platform a low centre of gravity and high static stability, properties that Song has argued are necessary for a robot expected to lift, push, and pull industrial loads safely in shared workspaces. Each arm is rated for a 5 kg payload at the wrist and up to 20 kg in proximal handling configurations. The base reaches a maximum speed of 1.9 m/s, equivalent to a brisk human walk.[1][3][13]
The robot is designed for round-the-clock factory operation and uses a hot-swappable battery system so that battery exchanges can be performed during a shift without interrupting tasks. Onboard inference runs on an NVIDIA Jetson Orin module, with cloud connectivity for fleet learning and remote teleoperation as a fallback control mode. Holiday Robotics describes Friday as offering both autonomous operation and a teleoperation mode, with control interfaces available through tablet, PC, voice, and gesture input.[3][13]
The most distinctive subsystem on Friday is its pair of in-house dexterous hands. Each hand carries 20 degrees of freedom across ten fingers (five per hand), weighs only about 500 grams, and is rated for up to 5 kg of load per individual finger joint. The actuators support backdrivability across all motion axes, allowing the hand to comply softly with external forces rather than fighting them, a property useful for delicate assembly and human-safe interaction.[1][6][14]
In place of the proprietary multi-thousand-dollar tactile arrays sold by some humanoid suppliers, Holiday Robotics designed its own magnetic tactile sensor that costs approximately 30,000 KRW (around 21 to 22 US dollars) per module to manufacture. The sensor uses a magnet embedded in a silicone surface; as external force deforms the silicone, the displacement of the magnet is read by a Hall-effect element underneath, which can resolve forces as low as roughly 0.01 newton (about 1 gram of contact pressure) and can also infer direction and torque. The sensors run at a 1,900 Hz refresh rate. Holiday Robotics covers approximately 80 percent of the hand's outer surface with these modules, including the entire palm and the inner surfaces of the fingers, rather than only the fingertips. Because the silicone wears with use, sensor pads are designed as replaceable modules, an approach Song has compared to trimming fingernails. The combination of full-surface tactile coverage and low per-sensor cost is central to the company's commercial pitch, since high-quality tactile feedback has historically been one of the most expensive line items on a dexterous humanoid hand.[1][13][15]
Song has repeatedly framed Friday's design around the assertion that "most of the value comes from the hands" in an industrial setting. The company explicitly de-emphasises bipedal walking on the grounds that legs add cost, complexity, weight, and failure modes without delivering proportional value in a typical factory aisle, where a wheeled chassis can cover the relevant ground. Resources saved on legs are redirected to high-degree-of-freedom hands, dense tactile sensing, and a reinforcement-learning control stack. The company has also designed for safety predictability by selecting from a library of vetted behaviours rather than letting an end-to-end neural policy directly command motors, an approach discussed below.[1][12]
Holiday Sim is the company's proprietary simulation environment, designed specifically to close the so-called "sim-to-real" gap that has historically plagued reinforcement-learning approaches to manipulation. The simulator focuses on physically accurate modelling of soft-body contact, deformable materials, and friction, areas where general-purpose robotics simulators often perform poorly and which are central to dexterous manipulation tasks such as inserting flexible cables, picking deformable parts, or grasping objects with unknown surface properties. Policies trained in Holiday Sim are then transferred to physical Friday hardware, where the company conducts iterative refinement.[1][6]
In place of the vision-language-action (VLA) models that several competing humanoid developers have adopted, in which a single neural network maps visual and language inputs directly to motor commands, Holiday Robotics has built a proprietary Vision-Language-Skill (VLS) framework. In the VLS architecture, a perception and language stack interprets the scene and the task, but instead of emitting raw motor torques it selects from a library of pre-verified low-level control "skills" such as grasp, insert, push, pour, or place. Each skill is implemented and tested independently and exposes well-defined preconditions and post-conditions.[1][3]
The company has described this as a "whitebox" approach in contrast to the largely opaque end-to-end policies emerging from VLA research. The advantages it claims are predictability and the ability to certify behaviour for industrial deployment, since each skill in the library can be inspected, tested, and bounded individually, and unsafe motions can be ruled out at the skill-selection layer rather than depending on emergent behaviour from a learned controller. The trade-off is that VLS will be less flexible than fully end-to-end policies for tasks that fall outside the existing skill library.[1]
For learning new skills the company favours simulation-based reinforcement learning over imitation learning from human demonstrations, the approach used in many recent VLA systems. Song has compared the strategy explicitly to DeepMind's AlphaGo Zero, which mastered Go without human game records by playing against itself: the goal is to let the robot accumulate experience through thousands of trial-and-error iterations in a virtual environment that approximates reality, rather than relying on labour-intensive teleoperated demonstrations. The company has cited reported figures from competitors (including descriptions of engineers at Tesla repeating motions up to 20,000 times to collect data for the Optimus humanoid) as evidence that the demonstration-driven approach scales poorly. Holiday Robotics positions its self-play training pipeline as a way to minimise the data collection and engineering effort required to teach a humanoid a new task.[6][9][12]
Holiday Robotics has positioned itself as a software-and-platform partner to existing system integrators rather than a direct supplier of factory automation services. Under this business-to-business model the company supplies Friday units and the underlying control software, while integrator partners handle deployment, line-side integration, and customer support. Song has said the company is currently conducting proofs of concept with major Korean manufacturers, although specific customer names had not been disclosed publicly as of the company's late-2025 announcements.[9][12]
The company's publicly stated production roadmap is:
The company has stated a target list price of approximately 100 million KRW (about $75,000) per unit at launch, a figure it justifies by reference to the annualised cost of a manufacturing worker. Song has been explicit that he views Holiday Robotics as a long-horizon, scale-driven business and has publicly favoured an eventual initial public offering rather than another early acquisition exit, arguing that the addressable market for humanoid labour is potentially larger than the combined automotive and smartphone industries.[9][13][15]
Holiday Robotics' August 2024 seed round drew attention in the Korean tech press for its size, especially against the backdrop of a broader funding slowdown for Korean startups in 2024. KED Global highlighted the round as one of several outliers showing that AI and robotics ventures continued to attract capital despite the wider funding drought, and Stonebridge Ventures' Choi Dong-youl publicly described the team's combined AI-software and hardware-controller expertise as the principal investment thesis.[4]
The October 2025 unveiling of Friday was covered by both Korean technology outlets and English-language humanoid trackers including Humanoids Daily, which characterised the design as a deliberate counter-position to bipedal-first humanoid programmes by Tesla, Figure AI, Agility Robotics, and Chinese rivals such as Unitree Robotics. Coverage focused in particular on the claim that Holiday Robotics had reduced the per-sensor cost of high-resolution tactile feedback to roughly $20 through its custom magnetic-Hall-effect design, an order-of-magnitude reduction from prevailing market prices for comparable sensors. Analysts cited by Korean outlets noted that price-aggressive industrial robot entrants from China, particularly Unitree, would be the company's most direct commercial competitors and that the wheeled-base, dexterous-hand design represented a coherent strategy for differentiating on cost and reliability rather than on athletic performance.[1][12]