At automotive factories in Germany and Hungary, humanoid robots have been tested on a task that requires more than simply following a set of repetitive instructions. The robots must identify components in supply bins, grasp them, and arrange them in the correct sequence for vehicle assembly.
In November 2026, Google DeepMind reported that its Gemini Robotics 2.5 system, working with Apptronik's Apollo 2 robots, achieved a 96% success rate in a proof-of-concept trial at Mercedes-Benz production facilities. The result demonstrated the system's ability to coordinate several physical actions in a real industrial environment, although it did not establish long-term reliability across an entire factory.
The experiment illustrates a broader transformation in computing. Machines are becoming increasingly capable of handling problems that require interpreting information, evaluating alternatives, and adapting their responses.
Some of these machines are physical robots. Others are powerful computers running artificial intelligence models that work on mathematical proofs, scientific research, weather forecasts, and industrial operations.
Their capabilities differ substantially, but they share an important underlying technology. Machine learning allows computer systems to acquire useful patterns from data rather than relying entirely on instructions written in advance.
How Machines Learn to Solve Problems
Traditional computer programs follow instructions designed by human programmers. They can perform extraordinarily complex calculations, but their behavior is determined by the algorithms and rules they have been given.
Machine learning introduces another approach.
Instead of specifying every rule needed to complete a task, developers train a model using data. During training, the model adjusts its internal parameters to improve its performance against a defined objective.
A system trained to recognize objects in photographs, for example, learns statistical relationships between visual features and the objects they represent. A weather prediction model learns patterns in atmospheric data, while a robotics model can learn how visual observations relate to physical actions.
Different applications require different learning methods.
Supervised learning uses examples with known outcomes. Reinforcement learning improves decision-making through feedback on actions or solutions. Other approaches train models to identify patterns and relationships within large collections of data.
Modern systems can also combine learned models with search algorithms, software tools, simulations, and verification procedures.
This combination is particularly important for complex problems. Producing a plausible answer is not the same as producing a correct one. Systems become more useful when their proposed solutions can be tested against mathematical rules, experimental observations, or measurable operating results.
Artificial Intelligence Is Advancing Mathematical Reasoning
Mathematics provides one of the clearest demonstrations of how machine learning is progressing beyond pattern recognition.
Difficult mathematical problems often require several connected steps, the ability to consider alternative approaches, and rigorous evidence that a proposed solution is correct.
In July 2025, Google DeepMind announced that an advanced version of Gemini with Deep Think had solved five of the six problems at the International Mathematical Olympiad.
The system received 35 out of 42 possible points, reaching the competition's gold-medal standard. Its solutions were officially graded and certified by Olympiad coordinators using the same criteria applied to student solutions.
The achievement represented an advance over Google DeepMind's 2024 systems, AlphaProof and AlphaGeometry 2, which had achieved a silver-medal-standard score of 28 points.
The distinction was not limited to the number of problems solved. The 2025 system generated mathematical proofs directly from natural-language problem descriptions within the competition time limit, whereas the earlier systems required specialized formal representations and considerably more computation.
Google DeepMind attributed part of the improvement to approaches that allow its model to explore multiple potential solutions and combine useful reasoning steps.
These results demonstrate substantial progress on a demanding mathematical evaluation. They do not, however, establish that artificial intelligence can solve arbitrary mathematical problems or independently conduct every stage of mathematical research.
The broader significance lies in the development of systems that can work through difficult chains of reasoning and produce solutions that human experts can evaluate. Google DeepMind's published results provide a concrete example of that capability.
Finding Better Solutions to Industrial Problems
Complex problem-solving extends well beyond mathematics.
Businesses regularly encounter optimization problems in which numerous possible decisions must be evaluated against practical constraints.
A logistics company must determine efficient delivery routes. A semiconductor designer must balance performance and physical constraints. A computing company must allocate processing capacity while controlling costs.
These problems can involve enormous numbers of possible combinations, making exhaustive evaluation impractical.
In May 2025, Google DeepMind introduced AlphaEvolve, a system designed to discover and improve algorithms by combining language models with automated evaluation.
Rather than simply generating code, AlphaEvolve proposes candidate solutions, measures their performance, and uses the results to guide further improvements.
The system has since been applied to computing infrastructure, semiconductor design, scientific research, and commercial optimization.
In a May 2026 update, Google DeepMind reported that logistics company FM Logistic used AlphaEvolve to improve routing efficiency by 10.4% compared with its previous heavily optimized solutions.
According to the reported results, the changes saved more than 15,000 kilometers of travel annually.
The percentage measures an improvement in the company's routing optimization, not a 10.4% reduction in its entire logistics budget.
In another application, AlphaEvolve helped refine data-management algorithms in Google's Spanner database, reducing write amplification by 20%. Write amplification measures how much data is physically written to storage relative to the amount originally requested.
Such results illustrate how artificial intelligence can contribute directly to operational improvements when a problem has a measurable objective.
The essential feature is verification. A proposed routing algorithm can be evaluated against distance and operational constraints. A computing algorithm can be tested for correctness and efficiency.
This allows machine-generated ideas to be assessed through evidence rather than accepted solely because an AI model recommends them.
Solving Problems in Biology and Scientific Research
Scientific research presents a different kind of challenge.
Researchers frequently work with systems containing large numbers of interacting variables, limited observations, and relationships that are difficult to establish through conventional calculations alone.
Machine learning can help identify patterns and generate predictions that researchers subsequently investigate.
One of the most significant applications has emerged in structural biology.
Proteins perform essential functions within living organisms, and their three-dimensional structures are closely connected to how they work. Understanding these structures and their interactions can be important for biological research and drug development.
Google DeepMind and Isomorphic Labs introduced AlphaFold 3 in 2024 to predict the structures of complexes containing proteins, DNA, RNA, and other molecules.
In a peer-reviewed study published in Nature, the researchers reported that AlphaFold 3 outperformed several existing specialized methods across important categories of molecular structure prediction.
The model uses a diffusion-based approach to generate predictions of molecular structures and interactions.
Its contribution is the ability to help researchers investigate molecular arrangements that might otherwise require substantial computational or experimental effort.
However, predicting a molecular structure is not equivalent to demonstrating how a molecule will behave under all biological conditions.
AlphaFold 3 has documented limitations, including difficulties representing molecular dynamics and certain structural interactions. Experimental validation remains important.
The distinction matters because scientific problem-solving involves more than generating an answer. Researchers must determine whether a prediction reflects the behavior of the physical system they are studying.
Machine Learning Is Changing Weather Forecasting
Weather prediction illustrates how machine learning can move from research into an established operational service.
Traditional numerical weather prediction relies heavily on physical equations describing atmospheric behavior. Powerful computers use these equations, together with observations of current conditions, to estimate how weather systems will evolve.
Machine learning systems take a different approach by learning relationships from large collections of historical and analyzed weather data.
On February 25, 2025, the European Centre for Medium-Range Weather Forecasts (ECMWF) began operating its Artificial Intelligence Forecasting System, known as AIFS, alongside its established physics-based Integrated Forecasting System.
ECMWF reported that the original operational AIFS delivered improvements on several forecast measures, including tropical cyclone tracks, while requiring substantially less energy to generate a forecast.
The center expanded the service with an operational ensemble model in July 2025 and upgraded both its deterministic and ensemble AIFS systems to version 2 on May 12, 2026.
The current AIFS system produces forecasts extending up to 15 days, with new forecasts generated four times daily.
That 15-day figure describes the forecast horizon, not a guarantee of accurate predictions throughout the period.
The significance of AIFS is its operational use within a major forecasting institution. Machine learning is no longer confined to experimental weather models. It is contributing to routinely produced forecasts alongside established physical modeling methods.
Rather than eliminating conventional numerical forecasting, the technology is becoming another tool for analyzing complex atmospheric conditions.
Robots Are Learning to Work in the Physical World
Software-based problem-solving and physical robotics share certain learning techniques, but their operating environments create different challenges.
A computer program may be able to test thousands of candidate solutions without affecting the physical world. A robot must deal with moving objects, uncertain positions, changing surroundings, and the consequences of incorrect actions.
Manipulating an unfamiliar object can require visual recognition, spatial reasoning, motion planning, and continuous adjustment.
The November 2026 Gemini Robotics 2.5 demonstration highlighted these difficulties.
Google DeepMind evaluated the system across 200 different tasks, including handling objects with tools, tying bows, and manipulating delicate materials. In its reported evaluations, the newer model achieved higher success rates than earlier baselines, although the improvements varied by task.
The company also tested the technology with Apptronik and Mercedes-Benz in automotive component handling.
These were developer-reported evaluations and proof-of-concept results, not evidence that the robots could independently perform every task required in a manufacturing facility.
Autonomous vehicles provide another example of machine learning interacting with the physical world at commercial scale.
Waymo's self-driving technology uses cameras, lidar, radar, and AI-based perception and decision-making systems to interpret road conditions and navigate traffic.
According to Waymo's published operating data, its vehicles accumulated 271.3 million rider-only miles through June 2026 across the U.S. operating areas included in its safety analysis.
Rider-only miles refer to driving conducted without a human driver behind the wheel. They measure actual autonomous driving distance rather than simulated mileage or the number of passenger trips.
Waymo also publishes comparisons between its crash rates and human-driver benchmarks adjusted for the areas in which it operates.
The results provide evidence about its performance under specific operating conditions. They should not be interpreted as proof that autonomous driving systems can safely handle every road, weather condition, or unusual situation.
Together, robotics and autonomous driving demonstrate both the potential and difficulty of moving machine learning from digital environments into physical operations.
Why Complex Problems Remain Difficult for AI
Despite rapid progress, artificial intelligence systems continue to face important limitations.
One challenge is generalization, or the ability to apply learned knowledge successfully to situations that differ from those encountered during training.
A model may perform exceptionally well on a mathematical benchmark yet struggle with an unfamiliar practical task. A robot may reliably handle specific components but perform less consistently when presented with different objects or arrangements.
Reliability becomes particularly important when a task requires many consecutive decisions.
The 2026 AI Index Report, published by Stanford University's Institute for Human-Centered Artificial Intelligence, documented substantial improvements in AI agents that operate computer interfaces.
On OSWorld, a benchmark that evaluates agents on tasks across computer operating systems, the report found that performance had increased from approximately 12% to 66.3%.
That result remained below the reported human baseline of 72.35%.
The benchmark measures task performance in a structured evaluation environment. It does not directly measure the percentage of entire jobs that artificial intelligence can perform or the reliability of every commercially available AI system.
Nevertheless, it illustrates an important distinction. Machines can make major improvements in solving complex tasks while still failing a meaningful share of attempts.
Other limitations include inaccurate outputs, sensitivity to changing conditions, difficulty evaluating uncertain information, and the computational resources required to train and operate advanced models.
These constraints explain why verification, testing, and human oversight remain important, particularly in scientific, medical, industrial, and safety-critical applications.
The Next Stage of Machine Problem-Solving
The development of machine learning is changing what computers can contribute to difficult work.
Systems that once focused primarily on identifying patterns and generating predictions are increasingly being combined with methods for planning, searching, testing, and refining possible solutions.
The applications are already diverse.
Mathematical AI models can produce advanced proofs. Scientific models can predict molecular structures. Optimization systems can improve industrial algorithms. Weather models can generate operational forecasts. Robots and autonomous vehicles can perform increasingly sophisticated tasks in physical environments.
These achievements are not directly comparable. A mathematical competition score, a forecast horizon, an improvement in routing efficiency, and a robot's task success rate measure fundamentally different capabilities.
What connects them is the growing ability of machines to use learned information to address problems that cannot be solved efficiently through simple, fixed instructions.
The most important advances may therefore come not from systems that generate more answers, but from those that can produce solutions whose correctness, efficiency, and reliability can be demonstrated.
That distinction will determine how far machine learning progresses from impressive technical demonstrations toward dependable tools for science, industry, and everyday life.




