Every day on the factory floor of BMW's Spartanburg, South Carolina plant, a humanoid robot loads sheet-metal parts into production fixtures. Standing roughly five and a half feet tall, with two arms, two legs, and a camera-equipped head, it works ten-hour shifts without breaks. Unlike the industrial robots that have populated car factories since the 1960s, it was not explicitly programmed for this task. It learned it.
The robot, known as Figure 02, spent much of 2025 at the Spartanburg plant and contributed to the production of tens of thousands of BMW X3 vehicles. It represents something the robotics industry has long promised but rarely delivered outside research laboratories: a general-purpose machine capable of operating in environments designed for humans, trained rather than hard-coded. The factory of the future, it appears, is arriving though not in the form many once imagined.
For most of industrial history, robots have been tools of extraordinary precision and extreme inflexibility. Industrial arms from firms such as FANUC and ABB execute predefined sequences with near-perfect repeatability, but even minor changes moving a component a few inches, altering lighting, adjusting tolerances can require costly re-engineering. This rigidity confined automation to high-volume, repetitive manufacturing. Warehouses, farms, construction sites, and small factories largely resisted it.
Artificial intelligence is changing this equation. The shift is not simply that robots are becoming "smarter", but that they are being designed to generalise. Instead of receiving explicit instructions for every contingency, machines are trained on vast datasets of human motion, simulated physics, and real-world demonstrations. The result is systems that can apply experience from one context to unfamiliar situations.
The brain behind the brawn
Traditional industrial robots operate on coordinates and sequences: pick at position X, rotate by Y degrees, place at Z. The logic is deterministic and brittle. The machines execute instructions flawlessly but possess no understanding of the task.
AI-powered robots operate differently. They rely on "foundation models": large neural networks trained on diverse data to acquire general capabilities. In March 2025, NVIDIA introduced Isaac GR00T N1. The system combines a slower reasoning module that interprets vision and language with a fast control module that generates real-time motor actions: an explicit attempt to mirror the separation between human deliberation and reflex.
Training such models requires enormous data volumes: recordings of humans performing tasks, simulated robot trajectories, and demonstrations from physical robots. NVIDIA reports that combining synthetic data with real-world demonstrations improved performance by roughly 40%, an important economic advantage given the high cost of collecting physical data.
Physical Intelligence, a startup valued at $5.6 billion after raising $600 million in late 2025, has demonstrated similar ideas. Its π0 model showed that a single AI system trained across multiple robot platforms could transfer skills to machines it had never seen before, a capability known as cross-embodiment transfer. In April 2025, the company released a system capable of cleaning homes in unfamiliar environments a long-standing challenge in robotics.
Foundation models have given robots perception and generalisation, but they do not, by themselves, solve a harder problem: how machines improve through interaction with the physical world. For that, robotics is rediscovering an older idea: reinforcement learning, now fused with large-scale pretrained models.
Skild AI, is among the most explicit proponents of this approach. Rather than treating robots as passive executors of learned behaviours, Skild frames them as adaptive agents that continuously refine their actions through trial, feedback, and reward. Its core system combines large multimodal models for perception and reasoning with reinforcement learning loops that operate both in simulation and on physical robots.
The distinction matters. Foundation models excel at recognising patterns and imitating demonstrations, but they struggle when objectives are underspecified or environments change in subtle ways. Reinforcement learning allows robots to optimise directly for outcomes speed, energy efficiency, success rates rather than merely reproducing examples. Skild's emphasis is not on one-shot general intelligence, but on competence accumulation: robots that become measurably better the longer they operate.
This philosophy aligns with the direction taken by Physical Intelligence in its latest π0.6 model. Earlier versions of π0 focused primarily on large-scale imitation learning across diverse embodiments. π0.6 extends this by tightly integrating reinforcement learning, allowing policies to be fine-tuned through interaction after pretraining. In effect, the model learns how to act from demonstrations, then learns how to improve by doing.
The technical shift is subtle but important. Reinforcement learning is no longer used to train robots from scratch – a strategy that proved brittle and expensive in the past—but as a refinement layer atop powerful pretrained representations. This dramatically reduces the amount of real-world data required while retaining the benefits of optimisation through experience. Tasks that require precise force control, long-horizon planning, or recovery from failure historically weak points for imitation-based systems show the largest gains.
Early demonstrations suggest that π0.6-style systems adapt more robustly to distribution shifts: variations in object placement, lighting, surface friction, or tool wear. Rather than requiring retraining or manual recalibration, the robot adjusts its behaviour online, guided by reward signals. This is particularly relevant for industrial and warehouse settings, where environments are nominally structured but never static.
The commercial implications are significant. Reinforcement learning closes part of the gap between laboratory performance and real-world reliability – the so-called sim-to-real problem – that has plagued robotics for decades. It also changes the economics of deployment. A robot that improves over time amortises its training cost across a longer operational lifespan, making higher upfront prices more palatable.
Yet reinforcement learning also reintroduces familiar risks. Poorly specified reward functions can produce unintended behaviours. Online learning raises safety concerns when robots operate near humans. As a result, most deployments constrain reinforcement learning tightly, limiting exploration and relying on human oversight. The industry has learned, painfully, that optimisation without guardrails can be as dangerous as rigidity.
What emerges from efforts like Skild AI and π0.6 is a more pragmatic synthesis: foundation models for perception and generalisation, reinforcement learning for adaptation and performance tuning, and conservative safety layers to bound behaviour. The result is not autonomous machines that "think" in a human sense, but systems that steadily close the gap between scripted automation and genuine physical intelligence.
The humanoid gamble
Alongside specialised automation, companies are betting on general-purpose humanoid robots. The logic is straightforward: the built world is designed for human bodies. Doors, stairs, tools, and workstations assume bipedal locomotion and human dexterity. A robot with a human form factor could, in principle, integrate into existing infrastructure without costly redesign.
This idea has attracted extraordinary capital. Figure AI raised over $1 billion in September 2025 at a valuation of $39 billion, up from $2.6 billion eighteen months earlier. Backers include NVIDIA, Microsoft, Intel, and Jeff Bezos. The company plans to manufacture humanoids at a facility called BotQ, initially targeting 12,000 units per year, with ambitions to reach six-figure production volumes within four years.
Others are moving aggressively. In October 2025, SoftBank agreed to acquire ABB's robotics division for $5.4 billion, bringing together industrial robotics expertise and SoftBank's investments in firms such as Boston Dynamics and Agility Robotics. Chinese manufacturers have advanced fastest in volume: UBTECH, Unitree, and AgiBot collectively shipped more than 10,000 humanoid robots in 2025. Unitree's G1, priced around $16,000 at scale, undercuts many Western competitors and has reportedly been profitable for several years.
The barriers ahead
Despite the momentum, obstacles remain substantial. Most humanoid robots operate for only two to four hours on a single battery charge. Costs remain high: Figure AI rents robots for roughly $1,000 per month, positioning them as premium tools rather than mass-market labour substitutes.
Reliability and safety are deeper concerns. Robots working alongside humans in unstructured environments face regulatory scrutiny. The European Union's AI Act, taking full effect in August 2026, classifies many robotics applications as high-risk, requiring extensive documentation and oversight. In the United States, regulation remains fragmented and largely voluntary.
Integration with existing systems presents its own difficulties. Most factories operate with equipment and software from multiple decades; connecting new robotic systems to ageing infrastructure often proves more expensive than the robots themselves.
Perhaps most significantly, the skills required to deploy and maintain AI-powered robots remain scarce. The manufacturing sector faces a structural labour shortage—roughly 500,000 unfilled positions monthly in America alone—that automation is meant to address. But automation itself requires trained personnel: technicians who can troubleshoot robotic systems, programmers who can customise AI models, engineers who can integrate new equipment with existing processes. Employers report skills gaps at a 17-year high.
A new industrial revolution?
Predictions of robotic transformation have repeatedly fallen short since the first Unimate arm appeared in 1961. Artificial intelligence itself has endured multiple cycles of hype and disappointment. Yet several forces suggest the current moment is different: genuine advances in robot learning, unprecedented capital investment, and mounting demographic pressure as working-age populations shrink.
The most plausible near-term outcome is not a sudden replacement of human labour, but gradual, uneven adoption. Warehouse, surgical, and agricultural robots will continue to spread. Humanoids will begin in controlled industrial settings before expanding cautiously. The process will be slower and messier than conference presentations imply.
For executives, the implications are clearer than they have been in years. Many forms of automation now offer demonstrable returns. The harder questions concern timing and strategy. Companies that wait for perfect general-purpose robots may fall behind competitors building experience today. Those that rush into immature systems risk wasting capital.
The safest conclusion is that success will belong to organisations that approach AI-powered robotics with neither blind enthusiasm nor reflexive scepticism, but disciplined analysis. The robots are becoming capable. Figuring out how and where to use them remains the real challenge.