Why Robots Can’t Fold Your Clothes Yet — And Why That’s About to Change — Week 1 (PreQuel) -Episode 2

Physical AI Field Notes — Episode 2 — Week 1 | From an Enterprise AI Architect’s Notebook

Ten years of building AI systems. Fortune 500 clients. Cloud-native architectures. Agentic workflows. LLMs before they were mainstream.

And yet, standing in my own laundry room last week, I realized something embarrassing: not a single robot on the planet can reliably fold my clothes.

That question — why not? — turned out to be the most important question I could have asked at the start of this journey into Physical AI.

The Boston Dynamics Illusion

We’ve all seen the videos. A humanoid robot doing backflips. A robot dog sprinting across a field. Atlas performing gymnastics that most humans couldn’t pull off.

It’s impressive. It’s viral. And it creates a dangerously misleading picture of where robotics actually stands.

Because here’s the reality: those same robots, for all their athletic elegance, cannot walk into your house and fold laundry. They cannot unload your dishwasher. They cannot pick up a dropped pen from an unfamiliar table.

The gap between “robot doing a choreographed backflip in a controlled lab” and “robot doing useful work in an unstructured human environment” is enormous. And understanding why that gap exists is the foundation of everything that’s happening in Physical AI right now.

60 Years of Classical Robotics — And Its Ceiling

For most of robotics history, the field was fundamentally a mechanical engineering problem.

Building a robot meant sitting down with pencil and paper — or a whiteboard covered in equations — and mathematically modeling every single joint, every link, every degree of freedom. There was a whole course in engineering school called Kinematics and Dynamics of Machines dedicated to this. You derived equations for forward kinematics (given the joint angles, where does the end effector end up?) and inverse kinematics (given where you want the end effector, what should the joint angles be?).

I used to wonder how anyone did this for a robot with 30 joints instead of 6. The honest answer: it became almost impossibly complex.

This is Explicit Modeling — you build a precise mathematical description of the robot’s rigid bodies, their interactions with the environment, and every physical constraint in the system. The pipeline looks like this:

Sense → Estimate State → Map → Plan → Control

For six decades, robotic engineers got remarkably good at this. Industrial arms in car factories. Surgical robots. Warehouse picking systems. All of these work because the environment is controlled, the objects are predictable, and the task is repeatable.

But here’s the ceiling they ran into.

Contact friction is almost impossible to model accurately. The moment a robot gripper touches a real-world object — especially a deformable one like a piece of fabric, a soft fruit, or a crumpled paper bag — the rigid body physics assumptions break down completely. You can model a steel bolt. You cannot accurately model a T-shirt.

The pipelines are brittle. Change the lighting. Move the object two inches. Introduce a new type of cup the robot has never seen. The entire carefully-engineered pipeline fails. Every new scenario requires a new engineering effort. There is no generalization.

The data is being ignored. Here’s the one that really got my attention from an enterprise architecture perspective: the entire classical robotics approach was developed in an era of data scarcity. Engineers had to model everything analytically because there was no alternative. But we live in 2026. There are now massive open robotics datasets being contributed by research institutions around the world. The OpenX dataset alone was assembled from 22 different robots, across 21 institutions, demonstrating 527 distinct skills. Classical methods have no mechanism to leverage any of this.

The 2010s: Reinforcement Learning Enters the Picture

When deep learning started its rise in the early 2010s, researchers saw an obvious application: use RL to let robots learn from experience rather than hand-coded equations.

And it worked — in controlled settings. Simulated environments. Narrow, well-defined tasks. The research was exciting and the papers were impressive.

But here’s the problem: this remained firmly in the domain of specialists. You didn’t see engineers on LinkedIn posting about the robots they’d built using RL. You didn’t see startups deploying RL-trained robots in real warehouses. The barrier to entry was enormous — you needed deep expertise in RL theory, simulation engineering, and hardware. It was a field for PhDs, not practitioners.

Around the same time, Behavioral Cloning started gaining traction as an alternative. Rather than learning from trial and error with rewards, BC learns from watching expert humans perform a task and imitating what they see. At its core it’s supervised learning — which made it more accessible. But it had its own limitations, which I covered in my last post.

Neither approach fully cracked the problem of generalizable, deployable robotics.

The Convergence That Changes Everything

Here’s where I think we actually are right now — and why I believe this specific moment in robotics history is unlike any that came before it.

Several powerful trajectories are converging at the same time:

• Large Language Models — systems that understand and reason about language at near-human level
• Vision-Language Models — systems that jointly understand images and language
• Reinforcement Learning — agents that learn from interaction and feedback
• Diffusion Models — generative systems that can model complex, multimodal distributions

In isolation, each of these is a significant technology. But there is one field where all of them can come together and produce something entirely new: robotics.

Think about what LLMs fundamentally changed in NLP. Before LLMs, if you wanted a machine to understand language in your specific domain, you had to collect your own data and train your own model from scratch. After LLMs, someone else has already done the foundational work. You fine-tune and deploy.

The question the robotics community is now asking — and actively answering — is: can we build foundation models for robotics the same way we built them for language?

The answer, increasingly, is yes.

Vision-Language-Action Models: The New Frontier

The result of this convergence has a name: Vision-Language-Action (VLA) models.

A VLA model takes in visual input (what the robot sees), language input (what task it’s been given), and outputs actions (what the robot should do). It’s essentially a foundation model with a physical action head attached — the same transformer architecture that powers GPT, but grounded in the physical world.

The Pi0 model by Physical Intelligence is one of the most talked-about examples. Trained on large amounts of diverse robot demonstration data, it can generalize across tasks and embodiments in ways that classical robotics and even early RL approaches simply couldn’t.

This is the destination the course is building toward. And understanding the full journey — from rigid body kinematics equations all the way to VLA models — is what makes the destination make sense.

What This Means If You’re Coming From Enterprise AI

Here’s my synthesis after two days of learning, filtered through a decade of enterprise AI experience:

Classical robotics failed to scale for the same reason rule-based expert systems failed to scale in software — the world is too complex, too variable, and too full of edge cases to hand-engineer every scenario.

The shift to learning-based approaches in robotics is the same transition we’ve already lived through in enterprise AI — from hand-crafted feature engineering to learned representations, from expert systems to foundation models. We’ve already seen this movie. We just didn’t realize the sequel was being filmed in a factory with robot arms.

The difference is that in robotics, the stakes of getting it wrong aren’t a wrong prediction in a dashboard. They’re a robot gripper that crushes something it shouldn’t. Which is exactly why the field took longer — and why getting it right now, with the tools and data available in 2026, actually matters.

What’s Coming Next

Next few weeks: We get into the actual mechanics of Imitation Learning — Behavioral Cloning from first principles — and start working with real joint angle data and the LeRobot library. And then the part I’ve been waiting for: hands-on with the SO-101 robot arm.

If you missed first post to this prequel— where a robot trying to pour latte art taught me the multi-modality problem — I’ll link it below.

The journey from enterprise automation to robot arms continues.

Physical AI Field Notes is a weekly series documenting my transition from enterprise AI/ML architecture into Physical AI and embodied robotics. Follow along on LinkedIn for daily notes from the learning curve.

Questions, pushback, or your own robotics story? Drop it in the comments.

Tags: #PhysicalAI #Robotics #RobotLearning #EmbodiedAI #VLA #MachineLearning #AIArchitect #CareerChange #FoundationModels #EnterpriseAI