There is genuine enthusiasm right now around deploying robotics in small and mid-scale manufacturing — including the factories doing prefab components for construction. I had multiple conversations on this topic since "dancing" robots took the prime on Youtube. The pitch is simple: labor is expensive, robots are getting cheaper, the math should work. And on the surface, it seems to.
But there's a gap in that logic. And it has significant implications for anyone evaluating automation as a near-term solution in this space.
When people see robotics headlines — humanoids dancing, wheeled robots navigating warehouses, arms assembling consumer electronics — they're mostly seeing Tier 1: capable, increasingly affordable machines optimized for repetitive, well-defined tasks in structured environments.
These robots are real, and they're genuinely useful in the right context. Entry-level humanoid platforms designed for research and light tasks — such as the Unitree G1 ($16,000 base) and EngineAI T800 ($25,000) — are now commercially available [1][2]. At $50,000 per unit operating 20 hours a day over a five-year lifespan, the effective cost per productive hour works out to roughly $3–$8 — well below human labor rates almost anywhere in the US [3]. But these are light-duty platforms. Their payload capacity, reliability under sustained load, and manipulation precision are suited for research and controlled logistics — not the physical demands of manufacturing or prefab.
The problem is that these robots cannot do most of what construction and prefab actually require.
The tasks that define small-scale manufacturing and prefab — adaptive assembly, material handling across variable inputs, fitting components with tight tolerances, working in partially built or irregular environments — belong to Tier 2: robots capable of complex, dexterous manipulation with meaningful load-bearing capacity and real-world reliability. Tier 2 looks very different economically. The robot hardware alone typically starts at $80,000–$200,000; once you add integration, safety systems, tooling, and commissioning, total system cost commonly reaches $150,000–$500,000 [4].
At that price point, the labor cost parallel reasserts itself sharply.
In California, direct manual labor in construction and light manufacturing spans a wide range depending on skill classification — from roughly $22–$26/hour for general laborers to $35–$46/hour for skilled tradespeople such as carpenters, MEP workers, and prefab assemblers [5][6]. These are direct-cost figures only: no benefits, no payroll taxes, no workers' compensation, no supervision overhead. Fully loaded, the real cost for skilled trades runs closer to $60–$80 per hour per worker [7].
That's a high bar to clear. But it's also a precise target. The question for any automation investment in this space isn't "are robots cheaper than humans in the abstract?" It's "can this specific robot, doing this specific task, at this reliability level, beat $60–$80 fully loaded — including the cost of downtime, integration, and the human still needed to supervise it?"
For Tier 1 robots, the answer is often yes — but only for tasks they can actually perform. For the complex manual operations that dominate prefab and construction manufacturing, those robots aren't yet in the running.
For Tier 2 robots, the hardware cost alone often erodes the labor arbitrage before you've accounted for anything else. A $200,000 robot with a 5-year horizon, operating single-shift (8 hours/day, 250 days/year), costs roughly $20/hour on hardware alone — before maintenance, integration amortization, or the supervision it still requires. Add those in, and you're back in the range of the worker you were trying to replace [4].
The gap between Tier 1 and Tier 2 capability is fundamentally a dexterity problem.
Most commercial robots today excel at force and speed in static environments — the same motion, repeated thousands of times, with minimal variation. This works well in automotive stamping plants, fulfillment centers, and electronics assembly lines. It breaks down in environments where the inputs vary, the geometry changes, and the task requires real-time physical judgment.
Human hands are extraordinarily capable instruments. Dexterous manipulation — picking up an unfamiliar object, sensing resistance, adjusting grip pressure mid-task, working around an obstacle — draws on decades of neuromuscular development that robots are only beginning to approximate. Two specific capabilities remain elusive at commercial scale:
Force sensing and feedback. Knowing how hard to push, pull, or grip without relying solely on vision. Without this, robots break components, miss tolerances, or fail to complete tasks that require pressure-sensitive contact.
Cross-task generalization. The ability to transfer a learned skill to a slightly different scenario without retraining. A robot that can drive a screw in one configuration often cannot drive a screw in a configuration two inches to the left. A human apprentice can.
The consensus among researchers and industry analysts puts reliable, broad dexterity in unstructured environments at 2028–2032. For semi-structured environments — like a prefab factory with more predictable layouts and standardized components — the timeline is shorter: targeted pilots are underway now, and commercial-scale deployments are plausible in the 2027–2028 window for specific task types [8][9].
The companies making the most credible progress on dexterity can be grouped by their approach: some are building the intelligence layer, some are solving the hardware problem, and a few are trying to do both.
Physical Intelligence (π) is arguably the most-watched name in robotics AI right now. Founded by five of the most cited researchers in robot learning and backed by $400M from Bezos, Khosla Ventures, and the OpenAI Fund [10], their thesis is that a single generalist foundation model — trained across enough tasks and robot embodiments — can generalize to new manipulation challenges the way large language models generalize across text. Their flagship model, π0, uses a flow-matching architecture for action generation and has demonstrated results on manipulation benchmarks spanning folding laundry, routing cables, assembling boxes, and packing food containers. For construction and prefab, the relevant question is whether that generalization holds when the inputs are rougher, the tolerances vary, and the environment isn't a lab.
Skild AI is attacking the same intelligence-layer problem with a more commercially aggressive strategy. Founded in Pittsburgh in 2023 and now valued at $14B after a $1.4B round led by Nvidia and SoftBank [11], Skild has already deployed its "Skild Brain" foundation model on Foxconn's assembly lines building Nvidia Blackwell GPU server racks — what the companies describe as the first public mass deployment of a generalized physical AI system in advanced manufacturing [12]. The strategic bet is a software-licensing model: one AI brain that runs on any hardware from any OEM, eliminating per-robot retraining. The data flywheel is central to the thesis — the more tasks the robots perform, the more real-world data the model collects, compounding capability over time.
Eka Robotics emerged from stealth in April 2026 with a sharply different thesis: that vision-language-action (VLA) models, dominant in the field right now, are fundamentally too indirect for contact-rich physical tasks. Co-founded by MIT professor Pulkit Agrawal and former DeepMind researcher Tuomas Haarnoja [13], Eka's Vision-Force-Action (VFA) model teaches robots mass, friction, and inertia through practice in high-fidelity simulation, then transfers those skills to real factory environments. The demo that drew attention: their robot decelerates approaching a light bulb, searches for it haptically, and screws it into a socket — contact-sensitive precision that most current systems cannot achieve. The goal, per Agrawal, isn't human-level dexterity. It's superhuman.
RLWRLD takes a similar force-first approach at the model level, treating torque, tactile feedback, and working memory as native data modalities rather than signals derived from vision. Headquartered in Seoul and backed by $41M in total seed funding — including strategic investors CJ Logistics, Lotte, and Hanwha — their model is trained directly inside live industrial operations rather than in lab settings, creating a proprietary real-world data advantage [14].
Origami Robotics attacks the problem from the hardware side. Out of YC's Winter 2026 batch, they've built a high-DOF robotic hand with in-joint direct-drive motors paired with a co-designed data-collection glove that matches the hand's kinematics exactly — eliminating the embodiment gap that causes most sim-to-real failures and allowing real-world manipulation data to be deployed directly to hardware without lossy translation. They've already sold hands to physical AI labs including Amazon [15]. Their core argument: you cannot train good dexterity models without hardware purpose-built to collect dexterity data, making hardware and model a co-design problem from the start.
What these companies share, despite their different approaches, is a common recognition: the intelligence layer and the physical layer cannot be built independently. The data you can collect is constrained by your hardware. The models you can train are constrained by your data. The path to cost-competitive dexterity runs through closing that loop in the real world — not in simulation — and the companies building the strongest real-world data flywheel earliest will be very difficult to displace.
For small and mid-scale manufacturers serving the construction space, the honest near-term picture is this: full automation of complex manual tasks is not imminent, and the economics of capable robots don't yet consistently beat fully-loaded labor costs for those tasks.
But prefab manufacturing is the most tractable entry point in the built environment. The work is more repetitive than on-site construction, the environment is more controlled, and the tolerances are defined by design rather than field conditions. For specific, high-volume operations within a prefab facility — panel cutting, component staging, repetitive fastening — the ROI case for current-generation automation is already real.
The strategic value isn't just in today's savings. It's in the data. Manufacturers that deploy early collect the proprietary real-world manipulation data that will train the next generation of models. The feedback loop between deployment and capability compounds over time, and the gap between early movers and late adopters widens with every cycle.
The dancing robots in the media are not the robots that will automate prefab. But the companies solving the dexterity problem — quietly, in factories, on real tasks — are building toward the moment when the economics finally converge.
That moment is coming. It is not here yet. And the distance between those two statements is where the real investment thesis lives.
Alexey Dubov | BuildTech VC
References
[1] Unitree G1 pricing — botinfo.ai/articles/unitree-g1 (May 2026)
[2] EngineAI T800 pricing — theresarobotforthat.com/blog/humanoid-robot-cost-roi-breakdown (Jan 2026)
[3] Robot cost-per-hour ROI analysis — blog.robozaps.com/b/roi-of-humanoid-robots (Mar 2026)
[4] Industrial robot total system cost — standardbots.com/blog/how-much-do-robots-cost (2026)
[5] California construction laborer wages — salary.com/research/salary/alternate/construction-laborer-salary/ca (Mar 2026)
[6] California construction skilled wages — glassdoor.com/Salaries/california-construction-salary (Feb 2026)
[7] Fully loaded US manufacturing labor cost — theresarobotforthat.com/blog/humanoid-robot-cost-roi-breakdown (Jan 2026)
[8] Dexterity timeline 2028–2032 — blog.robozaps.com/b/challenges-in-humanoid-robotics (Mar 2026)
[9] Semi-structured deployments 2027–2028 — IDTechEx via articsledge.com/post/ai-humanoid-robots (Jan 2026)
[10] Physical Intelligence funding — roboticscenter.ai/companies/physical-intelligence (Apr 2026)
[11] Skild AI $14B valuation — pittsburghstartupnews.substack.com (Jan 2026)
[12] Skild on Foxconn/Nvidia lines — technical.ly/entrepreneurship/pittsburgh-skild-ai-nvidia-foxconn (Mar 2026)
[13] Eka Robotics founding — indiatechreport.in/2026/04/30/eka-robotics (Apr 2026)
[14] RLWRLD $41M funding — globenewswire.com/news-release/2026/02/26/rlwrld (Feb 2026)
[15] Origami Robotics — ycombinator.com/companies/origami-robotics (2026)
