If you follow AI, you know the headlines are dominated by embodied intelligence, humanoids, and massive funding rounds. But why aren’t these robots actually everywhere yet? I’ve spent the last few months tracking the robotics sector, watching the money flow, the academic breakthroughs, and the slow, grinding reality of deployment. This industry sits on a thrilling fault line; one side holds world-changing capital and revolutionary AI models, while the other is stuck in the logistics of moving atoms in the real world. This tension, for example the colossal gap between the $21 billion invested and the actual number of robots running, is what we must understand right now. This analysis serves as my deep dive into the four critical bottlenecks that are slowing the future, explaining why, despite the delays, the arrival of truly general-purpose robots remains inevitable.
In 2024, robotics investment officially hit a record high of $21 billion, fueled by explosive AI breakthroughs and significant defense spending. Cross-embodiment AI firms alone vacuumed up $5 billion, representing a 150% surge from 2023. This capital influx stands in sharp contrast to deployment numbers: only 622,000 industrial robots were installed globally, a volume barely exceeding the number of Teslas sold in three months.
The Distribution Problem
GPT-4 was trained on trillions of tokens scraped from the internet. By comparison, the largest robotics datasets currently top out at maybe a few hundred thousand demonstrations. But the gap isn’t merely about volume; it’s about what Sergey Levine accurately termed “the internet of robot data” problem, as that internet simply doesn’t exist.
For the majority of robots and tasks, large, generalized datasets remain a fantasy. Unlike LLMs that can learn from passive observation of scraped text, robots require action-labeled data that maps not just what happened, but precisely what action to take. Worse still, that data demands perfect temporal alignment across multiple modalities; vision, force, proprioception, and gripper states must synchronize flawlessly at 30Hz or higher.
Consider the sheer scale: Generalist AI’s GEN-0 model, built on 270,000 hours of real-world manipulation data, is the current frontier. They are adding another 10,000 hours weekly. That sounds impressive, but it’s a drop in the bucket; it’s still orders of magnitude less than what powered the LLM revolution, and we need to be clear about that.
A single teleoperation session can easily generate gigabytes, but most of it is functionally noise unless you precisely capture the key moments, such as the subtle pre-grasp adjustments, the complex contact dynamics, and the critical recovery from slip.
Physical Intelligence’s π0 model is the most important evidence of a path forward. It learned from demonstrations across 30 different robot embodiments, totaling 900K trajectories. This cross-embodiment training is the unlock that allows a single policy to control vastly different hardware, from arms to quadrupeds. However, achieving that kind of data diversity demands either massive, bespoke infrastructure or truly clever data multiplication via simulation.
Two Paths Forward
I see the industry consolidating rapidly into two distinct camps, each optimizing for a different set of constraints.
The first camp, focused on narrow tasks like warehouse automation, is winning the deployment race. They use teleoperation and sheer scale to brute-force data collection, resulting in task-specific models that reliably hit 80-95% of human performance. Covariant, for example, built their model on “warehouse-scale” data from hundreds of connected robots using continuous fleet learning. It’s a successful strategy because the task distribution is contained and the environment is controlled, making the problem tractable.
The second camp, the Generalist policy labs, is betting on true flexibility. They need models that can walk into an unseen kitchen and cook breakfast, or enter a factory floor and instantly adapt to a novel task. This requires fundamentally different data strategies; learning from human videos to capture the full range of manipulation, training on internet-scale vision-language data to build deep semantic understanding, and leveraging massive simulation to generate variations of physical interactions.
In my view, the key architectural shift arrived in early 2025; the emergence of “two-brain” systems that cleanly separate slow, high-level reasoning from fast, low-level control. Physical Intelligence’s π0.5, DeepMind’s Gemini Robotics, and NVIDIA’s GR00T all leverage this architecture. It consists of a high-level VLM planner that decides what to do paired with a low-level controller running at 50+ Hz that handles the precise how to move.
This is why π0.5 can now handle a generalized task like “clear the table” in a completely novel, unseen environment. More impressively, if you interrupt mid-task and say “No pickles, I’m allergic,” it adapts its plan on the fly. This level of real-time, flexible adaptation was genuinely impossible just six months ago.
The Reaction Time Problem
The bottom line is that current systems are still too slow. Today’s frontier models operate at roughly 250ms between perception and action. That’s fine for picking up a stationary cup, but it’s a non-starter for true dexterous manipulation or, more importantly, safe human interaction.
Consumer-grade robotics—the robots we want in our homes—requires sub-50ms reaction times, what Eric Jang famously dubbed the “ultra instinct” threshold. Humans make kinematic decisions at roughly 10Hz. If a robot is expected to hand you a tool, fluidly mirror your gestures, or safely notice you approaching a door, it must react at similar speeds. The fastest deployed systems can hit 100Hz for highly specific industrial tasks, but achieving general-purpose, high-dexterity manipulation at that speed remains one of the field’s great unsolved problems.
This isn’t just a computational issue demanding faster models. It’s an architectural problem concerning action tokenization, which is the process of compressing smooth motion into discrete codes, on-device inference, or running the full model locally instead of in the cloud, and overall end-to-end system design. Recent, promising work like FAST and BEAST is demonstrating how smarter action encoding can improve both precision and latency, but the sub-50ms general manipulation threshold is, as of today, still out of reach.
The Simulation Question
Since real-world robot data is a prohibitively expensive bottleneck, every major lab is aggressively building synthetic data engines. The tradeoffs in approach are stark, however.
Physics simulators like ManiSkill3 are incredibly powerful. They can generate 30,000 frames per second by keeping the entire world on the GPU. They are fast and deterministic, but they struggle to capture the granular quirks of real-world physics. A robot trained purely in sim will stumble when it encounters unexpected surface friction or highly deformable objects.
World models, like Wayve’s GAIA-2 and DeepMind’s Genie 3, take an entirely different approach. They train neural networks to predict the future state when an action is taken. GAIA-2 can generate multi-camera driving scenarios, including systematically varying rare events. Genie 3 can create navigable 3D worlds from text prompts that persist for minutes at 24 fps.
The critical challenge with these models, however, is one of controllability. They can imagine plausible futures, but you cannot yet precisely specify, “apply exactly 5N of force at this precise angle.” Recent work like ASAP is attacking this problem differently; it involves training explicit error correction models that dynamically adapt policies to reality’s inherent quirks, rather than chasing a mathematically perfect simulation.
To my way of thinking, learning from human video is proving to be the most radically scalable approach. Physical Intelligence and other labs are training models on vast amounts of internet video, extracting latent action data from visual observation. This approach elegantly sidesteps the sim-to-real gap entirely, letting us learn from the true, global distribution of physical interactions. The limitation is that you still need real robot data for grounding the model, but the ratio is dramatically improving. Meta’s V-JEPA showed that you can pretrain on a million hours of internet video, then add just 62 hours of targeted robot interaction data and achieve surprisingly strong manipulation performance.
What’s Actually Deployed
The raw deployment numbers clearly reveal the gap between cutting-edge research and mass production.
Globally, 4.3 million industrial robots are in operation. The vast majority of these machines are concentrated in automotive and electronics manufacturing. By 2025, 16 million service robots will have been deployed, and 57% of those now leverage some form of AI for autonomy.
The year 2024 saw the first paying deployments of humanoid robots. Specifically, Agility’s Digit was deployed at a Spanx facility, while Figure 02 began shipping to BMW for essential automotive parts handling. Robotaxis from Waymo continue to scale, currently providing 150,000 trips weekly across four major US cities.
Context, however, is everything. South Korea, the undisputed leader, boasts 1,012 robots per 10,000 employees; that’s roughly one robot for every 71 human workers in manufacturing. The US trails significantly with 295 robots per 10,000 employees, and the global average sits at a mere 151.
The data confirms that most currently deployed robots still perform only narrow, highly repetitive tasks. The companies automating basic functions like warehouses, industrial sorting, and inventory management are the ones truly making money. Meanwhile, general-purpose manipulation remains largely confined to pilots and demonstrations.
Industry Structure Taking Shape
I see the industry consolidating rapidly into three distinct, critical layers.
First, the Labs training general foundation models, for example Physical Intelligence, Skild, DeepMind, and NVIDIA, require massive, continuous capital. For instance, Skild recently raised $300M in a Series A, and Physical Intelligence raised $400M. Only a handful of players can compete at this level. The barriers to entry are incredibly steep, demanding diverse robot fleets for data collection, specialized compute infrastructure, and unique teams that can expertly navigate both ML research and robotics systems engineering.
The second layer consists of companies deploying robots for specific tasks. These firms fine-tune the general models on domain-specific data, building their moats through customer relationships, deep operational expertise, and proprietary deployment datasets. The self-reinforcing loop here is clear; the foundation labs need diverse real-world data, and the deployment companies need access to better baseline models.
Finally, hardware manufacturers are fragmenting sharply by form factor. Humanoids, such as Figure, 1X, and Tesla Optimus, aggressively target general tasks in human environments. Collaborative robots, or Cobots, like Universal Robots, optimize for specific industrial workflows; they have over 100,000 units deployed. Specialized form factors, such as quadrupeds for inspection or snake robots for confined spaces, serve niches where mobility constraints dominate.
The unresolved business question is where the high margins will ultimately concentrate. Will the general model providers become the dominant platform layer, or will vertical operators who own the customer relationships and compliance workflows defend the higher profits? While the analogy to LLMs suggests the former, robotics has notoriously stickier and more complex integration requirements.
The Timeline Ahead
Extrapolating from the insights in Android Dreams and current research trajectories, I see the following timeline ahead.
2025-2026
The first true AI-controlled robots will replace humans in narrow verticals at significant scale. Task-specific robotics companies will deploy thousands of units and rapidly scale to hundreds of millions in revenue.
2026-2030
General models will improve dramatically through heterogeneous data collection and scaled reinforcement learning on deployed fleets. As a result, robots will expand beyond basic industrial picking into major sectors like manufacturing, farming, and construction. Critically, data collection methods will begin shifting away from manual teleoperation toward leveraging human video and efficient on-policy learning.
2030-2035
Humanoids will begin handling general tasks that require complex, long-horizon planning and real-time adaptation. The manufacturing growth loop will accelerate; robots will actively begin building robots. Concurrently, government subsidies and geopolitical competition will intensify around critical actuator supply chains and rare earth processing.
2035-2045
Embodied general intelligence, the true goal, will arrive. Robots will achieve human-level performance across most tasks that contribute to GDP. Online learning will scale through truly massive global deployment fleets, and the automation wave will decisively hit service work, healthcare, education, and our homes.
This aggressive timeline assumes continued, simultaneous progress on four major constraints; data collection scaling, robust sim-to-real transfer, achieving sub-50ms dexterity, and implementing long-term memory for multi-step tasks. It is important to remember that any one of these constraints could prove harder than currently expected.
What to Watch
These are the leading indicators I believe matter far more than flashy demos:
- Deployment numbers: How many robots are actually working in production? Not pilots, not demos, but machines earning money for paying customers at scale.
- Task diversity per deployment: Can one robot handle 10 different tasks, or just 10 variants of a single task? Generalization always matters more than narrow performance.
- Data collection throughput: How many hours of diverse, multi-embodiment robot data can the leading labs collect weekly? This remains the critical bottleneck determining model improvement rates.
- Latency improvements: Are systems truly moving from 250ms toward the 50ms threshold? Until we hit sub-50ms, dexterous, human-safe manipulation remains fundamentally out of reach.
- Sim-to-real success rate: What percentage of capabilities demonstrated in simulation work seamlessly on the first real-world deployment? This single metric determines how much our investment in synthetic data actually helps.
Why This Matters
Manufacturing, logistics, and agriculture employ 2.3 billion people globally and represent $40 trillion in GDP. Automating even a fraction will create brutal, winner-take-most dynamics.
Geopolitically, China dominates manufacturing and hardware supply chains with 73% of global robot deployments in Asia. They control 90% of rare earth refining and 98% of magnet production. The US maintains advantages in AI software and compute infrastructure. This sets up a strategic competition; can America’s AI lead translate to hardware dominance before China’s manufacturing scale compounds into an insurmountable advantage?
The social implications are just beginning. As automation scales, displaced workers will need new support systems. Political fights around UBI, profit distribution, and how to share automation gains are already emerging. The fertility crisis could worsen as companion androids become sophisticated. The concentration of economic power in a handful of robotics companies raises serious questions about democratic governance in an automated economy.
On the technical side, solving robotics pushes AI in critical directions; long-term memory, online learning, safety guarantees, real-time decision-making, and uncertainty quantification. These capabilities will inevitably ripple back into language models and other AI systems.
The Open Questions
Several fundamental uncertainties remain:
- Which form factor wins? Humanoids offer flexibility but suffer from complexity. Cobots have superior economics but limited scope. Specialized robots offer high performance but remain niche. The answer probably varies by application, but where will the volume truly concentrate?
- What’s the right training recipe? How much real robot data versus simulation versus human video is optimal? Which tasks benefit most from each? The empirical answers are still rapidly emerging as labs iterate.
- Can simulation close the gap? Or will real-world data always be the primary bottleneck? The ratio of synthetic to real data that enables strong performance is improving, but it’s unclear where it plateaus.
- When does the manufacturing growth loop hit? Once robots can reliably build robots, costs should fall exponentially. But how long until that loop closes? Current estimates range from 2030 to 2040.
- How do labor markets adjust? Past automation waves took decades to re-train and redistribute workers. AI and robotics are moving at a radically faster pace, and the social infrastructure to manage that rapid displacement simply doesn’t exist yet.
Resources for Going Deeper
- Technical Deep Dives:
- Robotics Deep Dive 2025 – Comprehensive overview of architectures, data, and industry dynamics.
- Android Dreams – Long-term scenario analysis through 2045.
- Eric Jang’s Ultra Instinct – Why latency is the critical bottleneck.
- Brad Porter’s This Business of Robotics Foundation Models – A necessary reality check on progress curves.
- Key Papers:
- π0, π0.5 from Physical Intelligence – Cross-embodiment foundation models.
- Gemini Robotics from DeepMind – Hierarchical VLM control.
- GEN-0 from Generalist AI – Scaling to 270K hours of real-world data.
- NVIDIA GR00T N1 – Humanoid foundation model.
- FAST and BEAST – Action tokenization for low-latency control.
- Dex1B – Billion-scale dexterity training.
- Industry Analysis:
- F-Prime Capital’s State of Robotics 2025 – Investment and market dynamics.
- Coatue’s Robotics Won’t Have a ChatGPT Moment.
- IFR’s World Robotics R&D Programs 2025 – Government funding strategies globally.
- Community Resources:
- Rohit Bandaru’s blog on vision-language-action models.
- Vintage Data’s industry analysis.
- Sergey Levine on Dwarkesh Podcast.
The field has moved past the question of whether general-purpose robots are possible. The path is clear, even if the timeline remains uncertain. Now we’re just engineering our way there—one demonstration, one deployment, one dataset at a time.