Robots That Think and Act Like Humans Are Here
Discover smart robots that learn to perform complex human-like tasks. Our advanced AI enables them to understand and act intelligently in the real world.
By Sci Herald Editorial Team ·
Based on research by Grid Robotics, Qingfeng Lan, Duo Li, Siyao Li, Yiheng Lyu, Shijian Lu, Zhenghao Qi, Shizun Wang, Xinchao Wang, Ming Yang, Zhiqin Yang, Zuhao Yang, Tao Yu, first posted to arXiv in November 2026 — view the original study
TL;DR: A new generation of smart robots can now learn and adapt to complex tasks in the real world, much like a human would. This shift, powered by advanced AI models like GPT-6 Astra, moves robots beyond simple, pre-programmed routines toward genuine problem-solving.
Key Facts
- A new paradigm called agentic robotics uses powerful AI models to allow robots to understand, reason, and act in open-ended environments.
- These systems operate in a closed loop: they observe, interpret intent, plan, act, and then verify the outcome before adapting their strategy.
- The core of this intelligence, dubbed Embodied System 2, combines a foundation model's reasoning with a physical 'harness' that manages real-world constraints.
- Researchers propose a new framework to evaluate these robots, measuring both their AI smarts and their real-world performance under pressure.
- The ultimate goal is sustained autonomy, where robots can safely take on broader responsibilities, from single tasks to fulfilling ongoing roles.
From Pre-Programmed Machines to Thinking Partners
"Within weeks, frontier models with agentic harnesses matched or exceeded task-specific policies on many embodied benchmarks," Qingfeng Lan, Duo Li, Siyao Li, Yiheng Lyu, Shijian Lu, Zhenghao Qi, Shizun Wang, Xinchao Wang, Ming Yang, Zhiqin Yang, Zuhao Yang, and Tao Yu write. Instead of training a robot for one specific job, we can now give it a general-purpose brain and let it figure out the steps itself.
What Are Smart Robots That Learn?
Smart robots that learn are machines that use a powerful AI 'brain' to understand a goal, reason through the steps to achieve it, and adapt their plan based on what happens in the real world.
The key is a continuous loop of action and verification. The robot doesn't just blindly follow a script. It observes its environment, interprets what a human wants, reasons out a plan, calls upon its built-in skills (like gripping or moving), and then—crucially—checks to see if its action worked. If something goes wrong, it can try again or change its approach entirely. This closed-loop operation is what makes these robots appear to truly think and learn from experience.
The Three Systems of a Thinking Robot
The researchers break down an agentic robot's intelligence into three layered systems, much like the human mind.
- System 2 (The Strategist): This is the robot's deliberative, slow-thinking core. It handles the high-level reasoning, interprets your intent, and makes the plan. It's realized by a massive foundation model (the AI brain) working with an "embodied harness"—software that acts as a manager, organizing information and ensuring the robot's actions are safe and observable.
- System 1 (The Skilled Worker): This system contains the robot's pre-learned skills and tools. These are the equivalent of muscle memory: a trained policy for grasping objects, a module for navigating a room, or a function for running code. System 2 calls upon these skills to get the job done.
- System 0 (The Reflexes): This is the robot's body and its instant, real-time control systems. It handles the millisecond-level adjustments needed for balance, force feedback, and avoiding immediate collisions. It's the autonomic nervous system of the robot.
A real-time event monitor acts as a lookout, constantly watching for changes—like a person suddenly walking into the path of the robot—and can interrupt the slow-thinking System 2 to trigger a lightning-fast safety reaction in System 0.
How Do You Test a Robot's Intelligence?
Evaluating a robot that can think for itself is very difficult. Beating a level in a video game simulation is one thing; performing a complex physical task safely and reliably is another. The paper highlights a major challenge: the research community lacks a standard way to measure the performance of these new robots.
The authors propose a new evaluation framework that tests a robot's abilities across nine key dimensions: But the real world is the true test. They can read a system's state directly (e.g., database query). 2. Their actions take effect immediately and perfectly. 3. Outcomes are automatically verifiable (e.g., a calculation is right or wrong). 4. Data is abundant, and environments are cheap to copy and reset.
A physical robot lives in a world where none of these are true. It can't directly "see" that a glass is full; it must infer it from camera images. Its actions, like lifting a bottle, have mechanical delays and can fail subtly. Verifying if it poured the right amount of water requires more sensors and reasoning. And you can't just reset a real kitchen with the click of a button after a robot spills something.
This means that while the reasoning power of AI models transfers from the digital to the physical world, almost everything about how that reasoning is executed must be rebuilt from the ground up to deal with messiness, delay, and consequence.
The Road Ahead: From Tasks to Roles
The final step is entrusting them with standing roles ("manage the kitchen each morning" - L4).
This path toward sustained autonomy is the ultimate goal. It won't be achieved by building a better pre-programmed machine, but by creating partners that can learn, adapt, and act responsibly in our unpredictable world.
The Bigger Picture
This work is part of a massive shift in robotics and artificial intelligence. For decades, the field has been dominated by creating specialized systems that could do one thing very well. The new frontier is generalizability—creating systems that can apply their intelligence broadly across a wide range of problems, much like humans do.
This research provides a crucial framework for understanding and measuring this progress. It connects to other advances in AI, such as models that learn to act by watching videos or that learn sophisticated object manipulation. The challenges of getting AI to reason reliably in the physical world also overlap with the development of better AI models for coding and tool-use, where the digital environment is a testing ground for complex problem-solving.
The open questions are vast. How do we ensure these robots are safe? How do they learn continually from their limited real-world experience? How do we build trust with machines that make their own decisions? The roadmap laid out by Grid Robotics is the first step toward answering these questions systematically.
What This Means for You
This isn't just lab science. The implications for everyday life are profound. In the near future, robots with this kind of adaptive intelligence could:
- In the home: Help with elder care, managing chores by understanding routines and adapting to changing needs.
- In manufacturing: Work alongside humans on complex assembly lines, troubleshooting issues on the fly.
- In healthcare: Assist in laboratories, reliably performing sensitive experiments that require precision and adaptation.
The key is reliability. These robots aren't just programmed for a task; they are built to understand their goal and figure out how to achieve it even when things go wrong. This makes them far more useful and versatile partners in our daily lives.
Paper, Code and Resources
- Paper: Envisioning Agentic Robotics: A New Era of Robotic Systems on arXiv (a preprint server; not yet peer reviewed)
- Project page: grid-robotics-official.github.io/agentic-robotics-roadmap
- Discussion: Hugging Face Papers — 66 community upvotes as of November 16, 2026
Frequently Asked Questions
How is this different from other robots?
Most current robots are built for specific, repetitive tasks in controlled environments (like a factory arm welding a car). These new agentic robots are designed for open-ended environments (like your home) and can understand a general goal, reason through the steps, and adapt their plan when they encounter the unexpected.
What is an embodied harness?
Think of it as the AI model's personal assistant for the physical world.
Source: Read the original study