Robotics and AI: What Current Robots Actually Can (and Can't) Do

I’ve been watching robot demo videos with a professional eye since my time at Samsung, where robotics was always looming as the next integration challenge. The Boston Dynamics Atlas doing parkour. Figure’s robot handling warehouse goods. Unitree’s dog-shaped robot navigating stairs.

Here’s what most of those videos don’t show: what happens when the environment is slightly different than expected. A box that’s a different weight. A floor that’s a bit more slippery. A doorknob that requires torque in an unexpected direction. That’s where the real frontier is — and it’s genuinely harder than anything the demos suggest.

Key Takeaways

• Current robots excel in structured, repetitive environments (warehouse logistics, automotive manufacturing, specific surgical tasks) and struggle severely in unstructured ones (home environments, disaster response, general-purpose assistance).
• The “manipulation problem” — reliably grasping and manipulating novel objects in cluttered environments — remains robotics’ hardest unsolved challenge.
• Sim-to-real transfer (training robots in simulation, deploying in the real world) is making progress but still requires significant real-world fine-tuning.
• Foundation models for robotics (RT-2, Pi-Zero) are showing early promise in generalist robot behavior, but deployment-ready generalist robots are years away from your home.
• Kids interested in robotics have more accessible entry points now than ever — from $50 Arduino-based builds to sophisticated FIRST Robotics teams.

What Robots Actually Do Well

The robots that work reliably — deployed at scale in 2026 — all share one characteristic: they operate in environments designed or controlled for them.

Automotive manufacturing: The most mature application. Welding robots, paint robots, assembly robots. These work because the car bodies are delivered in precise positions, the welding points are exactly where the robot expects them, and the environment doesn’t change. FANUC, KUKA, and ABB have deployed these systems in millions of units over decades.

Warehouse logistics: Amazon’s fulfillment centers use Kiva (now Amazon Robotics) mobile drive units to move shelving pods to human pickers. These robots navigate warehouse floors that are mapped, marked, and maintained for robot navigation. The robot brings the shelf to a human; the human picks the item. The difficult manipulation — picking arbitrary small objects — still requires a human hand.

Drone delivery and aerial robotics: Well-controlled airspace, GPS-guided navigation, limited manipulation requirement (the package just drops or is lowered). Amazon, Wing, and others have deployed commercial drone delivery in specific geographic markets.

Surgical robots: The da Vinci surgical system has performed millions of minimally invasive procedures. These are teleoperated — a surgeon controls robot arms with precision instruments — rather than autonomous. The robot provides precision and scale translation; the human provides judgment.

Boston Dynamics Spot: The quadruped robot is genuinely deployed for inspection tasks in industrial settings — oil refineries, construction sites, power plants. It navigates terrain and carries sensors. It does not operate autonomously in novel situations; it follows planned routes and sends data back to humans.

What Robots Still Fail At

The most informative list is not what robots can do — it’s what they can’t.

The unstructured home environment is robotics’ white whale. Pick up dirty laundry from a bedroom floor. Load a dishwasher with dishes in random positions. Fold a t-shirt and put it in a drawer. These are tasks every parent has nagged their kid about — and they’re genuinely hard for robots because the variability of positions, textures, weights, and configurations is enormous. No two piles of laundry are alike.

Dexterous manipulation of novel objects. A robot trained to pick up a specific style of box at Amazon might fail on a slightly different box. Humans generalize manipulation effortlessly — you can pick up any cup you’ve never seen before because you understand the concept of “cup” and how to handle an object of that shape and apparent weight. Robots learn the specific instances, not the general concept. The 2020 Amazon Picking Challenge results showed that state-of-the-art robot manipulation systems achieved roughly 65% accuracy on novel object picking — versus near-100% for trained human workers.

Contact-rich manipulation. Turning a key in a lock. Tightening a screw that requires feeling for the right torque. Plugging in a USB cable (even humans struggle with this). These tasks require sensing the forces involved and adapting in real time — something called haptic feedback and compliance control that’s much harder to automate than vision-based manipulation.

Generalization from limited experience. A human child who has seen one door can open most doors they’ll ever encounter. Teaching a robot to open “any door” — not a specific door — remains an open problem.

The Sim-to-Real Transfer Problem

One of the most important concepts in modern robotics AI is simulation training. Training a robot in the real world is slow and expensive — robots break, training takes real time, errors are costly. Simulation allows training millions of trials in hours. But what the robot learns in simulation must work in the real world — and the real world is always messier than the simulation.

This gap is called the sim-to-real transfer problem. Simulated physics is an approximation. Simulated surfaces don’t capture the true friction variability of real materials. Simulated lighting doesn’t match the infinite variation of real lighting. Robots trained purely in simulation often fail badly when deployed.

Solutions being researched include domain randomization (randomizing simulation parameters to make the policy robust to variation) and domain adaptation (fine-tuning on real-world data after simulation training). These help, but significant engineering effort remains required for each new deployment environment.

Foundation Models for Robotics: RT-2 and Pi-Zero

The most promising recent development applies the same ideas behind large language models to robotics. Google DeepMind’s RT-2 (Robotic Transformer 2, 2023) trained a vision-language model — the kind underlying image-understanding AI — on both internet data and robot demonstration data simultaneously. The result was a robot that could follow natural language instructions on novel tasks it had never seen, using the internet-scale knowledge baked into the vision-language model.

Physical Intelligence (Pi)‘s Pi-Zero (2024) extended this with a flow-matching policy architecture that better handles the fine motor control requirements of real manipulation tasks. Early demonstrations showed a robot folding laundry, wiping surfaces, and performing assembly tasks — with significantly better generalization than prior systems.

These are research demonstrations, not products. But the direction is clear: the path to general-purpose robots runs through large foundation models that encode broad world knowledge, adapted to robotic action.

What Kids Can Build at Home and in School

The accessible robotics landscape has expanded significantly.

Platform	Cost	Age Range	What Kids Learn
LEGO Mindstorms / SPIKE Prime	$350–$500	8–16	Sensing, actuators, basic programming
Arduino + servo kit	$50–$150	12+	Electronics, C++ programming, real hardware
Raspberry Pi robot kit	$100–$200	13+	Linux, Python, camera integration
FIRST Robotics (FRC)	Free (team-based)	14–18	Full engineering process, 6-week build season
Boston Dynamics Spot for Education	~$75,000	University	Professional-grade platform
Open-source quadruped (Mini Pupper, etc.)	$400–$800	14+	Advanced electronics, ROS framework

FIRST Robotics Competition is underrated as a path into professional robotics. Alumni include engineers at Boston Dynamics, NASA, and every major robotics company. The 6-week build season to competition deadline mimics real engineering crunch cycles in a way that college coursework rarely does.

How to Teach Your Kid About Robotics and AI

Ages 5–8: Robot that can’t think

Build a simple robot that follows a fixed path — a LEGO wheeled robot with a preprogrammed sequence. Then give it a task it wasn’t programmed for. When it fails, ask: “What would the robot need to know to handle this? How would you teach it?” This makes the limitation of fixed programs concrete, and opens the question of what “learning” would mean for a machine.

Ages 9–12: Sensor fusion challenge

Give your child a blindfold and have them navigate a familiar room using only touch and sound. Then discuss: “A robot navigating a new room has to combine camera images, LiDAR (laser ranging), and force sensors to understand what’s around it. Why is combining multiple sensors better than using just one?” This introduces the concept of sensor fusion, which is foundational to robotics. Build a simple maze-following robot with an ultrasonic distance sensor to make it concrete.

Ages 13+: Try Robot Operating System (ROS) in simulation

ROS (Robot Operating System) is the standard middleware for robotics development. Gazebo is a free simulator that integrates with ROS. A teenager interested in robotics can install ROS on Ubuntu (or use a Docker container on any OS), load a simulated robot, and write a simple navigation script. The process is not trivial — ROS has a steep learning curve — but it’s the actual tool the field uses, and getting a simulated robot to do anything at all is genuinely satisfying. Learning to code is the prerequisite here.

The question to ask: “If you had to teach a robot to make a peanut butter sandwich, how many separate steps would you need to specify? What would it need to sense at each step? What could go wrong?”

What to Watch For Over the Next 3 Months

Month 1: Watch one Boston Dynamics demo video and one video of a robot failing at a task (there are many honest “robot fails” compilation videos from DARPA challenges and other sources). The contrast is instructive. Demos are selected for success. Engineering is defined by understanding failure modes.

Month 2: If your child is interested in robotics, look for a local FIRST Robotics team. Most teams welcome observers at practice sessions and competitions. The culture of collaboration, mentorship, and engineering problem-solving under real time pressure is something that can’t be replicated by watching YouTube.

Month 3: Follow the development of Pi-Zero and similar foundation model robotics systems. Physical Intelligence, 1X Technologies, and Apptronik are the companies closest to general-purpose manipulation capability. Their research publications and demo videos give a realistic picture of where the field actually is, as opposed to where breathless headlines claim it is.

Frequently Asked Questions

Are humanoid robots like Figure and 1X going to be in homes soon?

Not in the next 3–5 years in any meaningful sense. The manipulation challenges, the safety requirements for robots operating around children and pets, and the cost of the hardware all push this timeline out. The first commercial deployments will be in controlled industrial environments — logistics, manufacturing, inspection — where the environment can be adapted to the robot’s capabilities.

How do robots “see” the world?

Most robots use cameras (RGB cameras, depth cameras that use infrared to measure distance, or event cameras that detect changes rather than full frames), LiDAR (laser rangefinders that build 3D point clouds of the environment), and ultrasonic sensors for close-range detection. These signals are processed by AI models that classify objects, estimate poses, and build maps. The challenge is that none of these sensors are as robust as the human visual system in terms of handling lighting variation, rain, and novel situations.

What’s the difference between a robot and an AI?

A robot is a physical system that senses and acts in the world — it has sensors and actuators. AI is the software that controls what the robot does with those sensors and actuators. Not all robots have AI (many industrial robots run fixed programs), and AI doesn’t require a physical robot body. Robotics AI specifically refers to AI systems designed to control physical robots in dynamic environments.

My kid wants to study robotics. What major should they choose?

Robotics programs are increasingly available as standalone degrees at universities including Carnegie Mellon, Michigan, Georgia Tech, MIT, and others. Robotics crosses mechanical engineering (robot mechanics), electrical engineering (sensors and actuators), and computer science (AI and software). A robotics major addresses all three; alternatively, an EE or ME major with a CS minor covers most of the ground. The practical edge goes to students who build real things — competition robots, independent projects — alongside coursework.

---

About the author Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Brohan, A., et al. (2023). “RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control.” arXiv. https://arxiv.org/abs/2307.15818
2. Mahler, J., et al. (2017). “Dex-Net 2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds and Analytic Grasp Metrics.” Robotics: Science and Systems. https://arxiv.org/abs/1703.09312
3. Tobin, J., et al. (2017). “Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World.” IROS 2017. https://arxiv.org/abs/1703.06907
4. Black, K., et al. (2024). “π0: A Vision-Language-Action Flow Model for General Robot Control.” Physical Intelligence Technical Report. https://arxiv.org/abs/2410.24164
5. FIRST Robotics Competition. (2024). “Program Statistics.” https://www.firstinspires.org/robotics/frc
6. National Science Foundation. (2023). “National Robotics Initiative 3.0 Program.” https://www.nsf.gov/funding/pgm_summ.jsp?pims_id=503641
7. Zeng, A., et al. (2020). “Transporter Networks: Rearranging the Visual World for Robotic Manipulation.” CoRL 2020. https://arxiv.org/abs/2010.14406