Industrial Robots: $16.8b (’22) → $41.0 (’30) (CAGR 12.3%) By applications manufacturing (electronics 28%, automotive 25%), metal welding/cutting (20%?) By geography Robot installations by country (’22): China 50%, Japan 8%, US 7%, Korea 6%, Germany 5% Global: dominated by EU/Japan players. China: still dominated by EU/Japan. Chinese players M/S 15% Korea: HD현대 grew out of EU/Japan’s system integrator → M/S 20% Assessment huge, not growing fast enough bespoke software & hardware. not extensible much. bigger project in general (ASP ~$100k) dominated by major players from Japan & EU low market multiple in general. (PER 10x~20x PSR 1x~2x) not much venture-scale opportunities in traditional playbook Collaborative Robots (cobots): $2.14b (’24) → $11.1b (’30) (CAGR 31.6%) By applications not specified. but must be similar to traditional industrial robots (electronics, automative, metal handling (welding) as key 3 applications) (ASP $20k~50k) Assessment relatively small (~10%), fast growing (31% vs. 12%) different key players: Universal Robots, Doosan Robots … relative (much) higher market multiple (PER 20x~30x) (Doosan PSR 50x) some possibility of venture-scale return opps (e.g. Doosan Robotics $3.2b PSR 50x, Rainbow Robotics $1.8b PSR 100x), which are focused on the high-growth cobots segments (Not sure if this trend would be sustained in the long term, based on just 2 data points, but it seems bit of hype and other stories) Investment Thesis in the current industry Traditional Industrial Robots: Pain points: high initial cost, lack of flexibility, safety concerns Stable but slower growth → lower market multiples than cobots Thus, not considered a space for venture-scale return in its current form Cobots: Still smaller than traditional industrial robots, but much higher growth rate → higher market multiple at 2x~3x of industrial robots Opportunities in SMEs and industries seeking flexible automation solutions Overall observations Industrial robots 1/2 vs. Service robots 1/2: the service robots market grows faster than the industrial robots. For industrial robots by geography, Asia Pacific market is about 65% with China ~=50%, Japan 8%, Korea 6%. The global industrial robot market (including China) is dominated by players from Japan, EU (with Korean players with some market share in Korea). Chinese industrial robot suppliers are emerging, but yet have small market share in both China market (M/S 15%) and outside-China global market (M/S <5%). Chinese players in cobots market is bigger with 20~30% M/S, but the cobots market itself is relatively small at just about 10% of the industrial robot market yet. Korean industrial robot market is about 6% of the global market, and local Korean suppliers have the 30~40% market share in Korea, which translates as 2~3% global market share Korean manufacturing (as TAM of industrial robots): $350b (electronics $63b, automotive $138b, machinery $132b, ship building $24b) (주요산업동향지표 (KIET)) still, the automation rate is relatively low (30% in Samsung) largely due to lack of AI intelligence yet, which may present venture-scale opps with AI Industrial Robots Market

Professional service robots: $16.3b (’22) → $62.3b (’30) (CAGR 18.4%) 1. Medical Robots: $8 billion. CAGR 20% Utilized in surgeries, rehabilitation, and diagnostics 2. Logistics Robots: $6 billion, CAGR 25% Employed in warehouses and distribution centers for material handling 3. Agricultural Robots: $4 billion. CAGR 24% Assist in tasks like planting, harvesting, and monitoring crops. 4. Inspection and Maintenance Robots: $2 billion. CAGR 18% Used for infrastructure inspection and maintenance tasks. 5. Defense and Security Robots: $5 billion. CAGR 15% Deployed for surveillance, bomb disposal, etc Personal service robots: $39.7b (’23) → $267b (’33) (CAGR 21%) 1. Household Robots: $7 billion. CAGR 17% Include vacuum cleaners, lawn mowers, and window cleaners 2. Entertainment and Leisure Robots: $1.5 billion. CAGR 15% Robotic toys 3. Educational Robots: $1 billion. CAGR 19% educational robots for learning purposes 4. Assistive Robots: $0.5 billion. CAGR 25% Aid individuals with disabilities or the elderly in daily activities Key takeaways Much larger than industrial robots segment in aggregate, but, highly fragmented by verticals. Each vertical has very different requirements and tech, given that the top 5 players and the investment theses in respective categories are all different. When we need to evaluate investment opportunities in the services robots, we need to consider each verticals as separate markets, unless the target company is providing the all-encompassing underlying infra/ tech, e.g. LLM common for all verticals Service Robots Market

1. Foundation Model State of Market Leading researches/products led by Google etc. proliferates in the VLA (vision-language-action) Foundation Model for robotics Robotics Foundation Model (VLM/VLA models) Big Tech Google RT-X RT-2-X: closed source. model: 55B RT-2 + robotic data mixture (500 skills, 150k tasks, 1m+ workflows) RT-2 (’23.7): based on 알 수 없는 링크 & PaLM-E. RT-1 (’22.12): based on VC-1. Transformer-architecture RT-X: RTX-1, RTX-2 Pytorch implementation of the models RT-1-X and RT-2-X from the paper: "Open X-Embodiment: Robotic Learning Datasets and RT-X Models" RTX-1: takes in text and videos RTX-2: takes in images and text and interleaves them to form multi-modal sentences and outputs text tokens not a 7 dimensional vector of x,y,z,roll,pitch,yaw,and gripper Dataset: RTX-2 dataset on huggingface NVIDIA GEAR (Generalist Embodied Agent Research) Project GR00T: humanoid foundation model MimicPlay: imitation learning method Meta OK-Robot Open-source, research OpenVLA (’24.6) (Stanford) Octo (’24.5) (UC Berkeley) TinyVLA (’24.9) (China) 3D VLA (’24.3) (UMass, etc) Dataset: OpenX dataset (’23.10) New emerging robotics (VLA) Foundation Model startups Physical Intelligence Founded by ex-Google Research team Fuding: $70m Seed (’24.3. $400m post), and $400m Series A (’24.11. $2.4b post) LLM: π0 model. Dataset: N/A Tech overview: https://www.physicalintelligence.company/blog/pi0 Investors: Thrive, Lux, Jeff Bezos (Lead, Series A), Khosla, Squoia, Lux (Seed) Skild Founded by 2 IIT graduates/CMU Funding: $300m Series A (’24.7. $1.5b post) (Seed funding not disclosed) LLM: Skild Brain robotics Foundation Model Dataset: N/A Robots supported: resilient quadruped, vision-based humanoid Investors: Cuatue, Lightspeed, Jeff Bezos (Lead, Series A) World Labs Founded by Fei-Fei Li, a leading AI research and professor of Stanford Univeristy Funding: Seed (’24.4. details not disclosed), and $230m Series A (’24.9. over $1b post) LLM: Large World Model (LWM) for “spatial intelligence” Initially, 3D world generation “without limits – creating and editing virtual spaces complete with physics, semantics, and control” As the next steps, “World Labs seeks to enhance applications in areas such as robotics, augmented reality, and virtual reality. This advancement would enable AI systems to navigate and manipulate real-world spaces more effectively, leading to more intuitive and human-like interactions. For instance, robots could perform tasks with greater precision, and virtual environments could become more immersive and responsive.” (No further details available) Investors: Radical Ventures (Lead), NVentures, Intel Capital, AMD Ventures, Databricks Ventures (Series A) yet to be seen how the product & business strategy would unfold over time. Challenges Training data: high-cost, time-consuming (moat) Integration of intelligence + precise robot control (moat) end-to-end VLA models → low precisions yet (e.g. Figure-1 in BMW factory, PI-0 for laundry) integration of intelligence model (VLA, Diffusion Model) + robot control model (RL-based) Intelligence & precision achievable for production environments yet to be seen. when, how much? Handling diverse form factor: quaruped (robot dogs), biped (humanoids), dual arms (torso, double robot arms), wheeled/fixed Assessment Still in very early-stage; playbook yet to be developed (including product, GTM, pricing model) Failure of Covariant → most might go 'vertically-integrated' or 'stronger partnership with hardware' paths for product strategy Then, the pricing/biz model for the FM play would be a challenge → "Sell work, not tool" model might work? 2. Robot Hardware Key players Humanoids (torso/full-body, heeled/biped) Tesla Optimus, Figure, 1X Technologies, Agility, Apptronik, Boston Dynamics, Fourier Intelligence, Sanctuary, Unitree, XPENG (20-30 projects in US & China) Summary Quadruped (robot dogs): Boston Robotics … Dual arms (torso, double robot arms, wheeled/fixed): … Assessment Most players with vertically-integrated approach (hardware + software/FM) It is yet to be seen how intelligent these robots can develop with such vertically-integrated approach. Of course, there might emerge partnership with FM-only players (PI, Skild for now) as needed Most of the humanoids are still in early demo phase, and their real-world deployment is at least a few (2-3) years away. (ref: demo of Figure-1 at BMW factory, π0 for laundry folding) 3. Pre-training Data State of Market So far in last few years since trying to apply Gen AI to robotics, the pre-training data collection for robotics FM has been done ad hoc Key pain points quality & quantity of data diversity in action scenarios & environments high cost of data acquisitions & generations Primary methods Tele-operation: time-consuming, costly, limited actions Video analysis: video → 3d skeleton/action sequences → training for text-to-action ad: vast amount of training data disad: lack of precision in action sequences Just the beginning: e.g. Cognitive Intelligence Simulated data gen: Panoptics, Google RT-2 extract 3D motion/skeleton data from sensors/cameras in artificially created environments ad: cost-efficient (Panoptics) disad: as costly as tele-operation (Google RT-2) Startups Assessment The market size yet to be seen, esp. whether the huge demand for pre-training data is to emerge like the Scale AI case (for autonomous driving), and how many players can survive (few dominant players other than Scale AI) Longer GTM cycle expected, considering the value-chain: robots deployed with hardware & FM integrated ← hardware only, FM software only ← FM software ← pre-training data(Remember, Covariant?) 4. Others Reinforcement Learning Human-Robot Interactions (HRI) Multi-robot Coordination Sim-to-Real Transfer Edge AI Deployment Assessment Each area is considered very specific niche, highly technical, part of tech stack by verticals (e.g. multi-robot coord (aka. drone swarming) for defense or entertainment), might be of limited market size in themselves; Suspect if there are venture-scale opps in each area Opportunities in AI + Robotics