DROIDS Newsletter

The Helix 02 video draws you in from the beginning with the sound. The gears moving as Helix approaches the dishwasher. There is no background music. No excessive production value. It is four minutes of Helix unloading a dishwasher and putting the dishes away, closing a cabinet, unloading the cups (which are plastic, hold onto that detail because it is important), stacking them and putting them away in an upper cabinet, retrieving a mug and a cup at the same time, and putting both away. All of these movements are done autonomously.

Helix opens up a drawer to the side of the sink, and I was genuinely curious as to what it was doing. What’s in the drawer? Ah, a dishwasher pod. I was hooked on this video like it was a sci-fi thriller.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit droids.substack.com

NVIDIA just lobbed a heat-seeking missile at Tesla’s Full Self-Driving empire. It’s called Alpamayo, and it doesn’t just want to drive your car. It wants to reason through traffic like a real human—with verbal explanations, open-source weights, and a brain that learns from simulation. And it’s not a concept. It ships this year, in the new Mercedes-Benz CLA.

As NVIDIA summed up the problem: AVs must safely operate across an enormous range of driving conditions. Rare, complex scenarios, often called the “long tail,” remain some of the toughest challenges for autonomous systems to safely master. Traditional AV architectures separate perception and planning, which can limit scalability when new or unusual situations arise.

The Alpamayo Stack: Think Before You Steer

Named after a Peruvian peak, Alpamayo is NVIDIA’s new “reasoning AI” platform for cars: a portfolio of ~10-billion-parameter models that take in real-time video and sensor data, break it into decision sub-tasks, and generate a trajectory—all while explaining their reasoning in natural language.

Think ChatGPT, but for merging onto I-405. And then doing it again. And again. Safely. Verbally. Auditable. That’s Alpamayo’s hook: explainability.

It’s not just a model, either. It’s a stack:

* AlpaSim: High-fidelity closed-loop simulation that trains and tests AI agents in virtual cities.

* Halos: A backup “classical” safety stack that monitors the AI’s choices and can yank the wheel if needed.

* Open datasets: Released publicly, hosted on Hugging Face. Yes, the weights are coming too.

Mercedes Is First, But Not Last

The first commercial deployment is the 2026 Mercedes-Benz CLA, built on the MB.OS architecture and NVIDIA’s full DRIVE AV stack. What you get: a Level 2++ system that handles point-to-point driving in urban and highway settings—hands on, eyes up, but AI in the loop the whole way.

Behind the scenes, that stack combines Alpamayo’s neural-policy brain with classic safety logic. It watches, reasons, explains. And if it screws up, the backup stack steps in.

That car will be on U.S. roads in 2026. Europe follows. Asia comes in 2027. No overpromising. Just a slow, deliberate ramp with regulators and human drivers fully in the mix.

“

For more than a decade, the humanoid robot Atlas has served as a symbol of what advanced robotics might one day achieve. Developed by the Massachusetts-based firm Boston Dynamics, the robot became widely known for demonstrations of agility and balance that circulated online, even as its commercial future remained uncertain.

On Monday at the Consumer Electronics Show in Las Vegas, Boston Dynamics and its majority owner, Hyundai Motor Group, said that Atlas is now being prepared for real industrial use. Hyundai plans to begin deploying the humanoid robot in its factories starting in 2028, marking a shift from research demonstrations to commercial operation.

The announcement places Boston Dynamics among a growing group of companies attempting to introduce humanoid robots into workplaces, where expectations for reliability and safety are significantly higher than in laboratory settings.

From Research Platform to Product

Boston Dynamics was founded in 1992 as a spinout from the Massachusetts Institute of Technology and spent much of its early history focused on government-funded research into robotic mobility. The company changed hands several times before Hyundai acquired an 80 percent controlling stake in 2021.

Since then, Boston Dynamics has increasingly emphasized commercialization. It already sells Spot, a four-legged inspection robot, and Stretch, a mobile robot designed for warehouse unloading. Atlas, however, remained a research platform until now.

At CES, the company confirmed that the current version of Atlas is considered a commercial product. This version is fully electric, replacing earlier hydraulic systems, and is designed for repeatable manufacturing and long-term use in industrial environments.

F
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit droids.substack.com

I did not plan to stop in my tracks at the Humanoids Summit. But then a man wearing four robotic arms started high-fiving strangers. Summit goers stopped in awe, and he received rockstar treatment. People kept asking for selfies.

The company is Psyonic. The founder is Aadeel Akhtar. What he was wearing was a costume, but it functioned as a wearable demonstration of a serious idea about how robots will eventually learn to use their hands.

Akhtar calls it his Doc Ock suit, and it is an outstanding replica of the Marvel character. But, even if you don’t know anything about Doc Ock, you’d be hard-pressed not to be drawn in by the suit. Four robotic hands moved with intent, controlled by his own hands through gloves. Right hand controlling the right side. Left hand controlling the left. The setup was intuitive enough that you understood it before it was explained.

That is when the conversations started.

Psyonic builds advanced bionic hands that are used by both humans and robots. This is not a theoretical claim. Nearly 300 patients use the Ability Hand as a prosthetic, covered by insurance and Medicare. At the same time, more than 50 robotics companies use the exact same hand on robots, including NASA, Meta, Google, Amazon, and automotive manufacturers working through humanoid platforms.

That overlap is the point.

When I asked Akhtar where he sees the industry going, he did not talk about better motors or more parameters. He said the hand problem is a data problem. Robots struggle with dexterous manipulation not because we do not know how to build hands, but because we do not have the right training data. Especially when it comes to soft, deformable objects that require constant force adjustment.

A human knows how to pick up a fragile object without crushing it. Not by calculation, but by feel.

What makes Psyonic different is that the same hand goes on a human and on a robot. Human users are already doing the tasks robots are trying to learn. Picking and placing. Sorting. Folding laundry. Cooking. Working in industrial environments. Because the hardware is identical, the data transfers cleanly. This is real to real learning without translation loss.

That distinction matters more than it sounds.

Most robotics demos show hands mounted to tables, performing narrow tasks under ideal conditions. Psyonic’s hands live on real bodies. They experience friction, hesitation, mistakes, and recovery. They generate data shaped by the messiness of the real world. That data is what robots actually need.

In 2024, Psyonic won a rare three-shark deal on Shark Tank. It was a brief moment in the spotlight for a company that spends most of its time solving hard problems far from TV cameras.

Seeing a PYSONIC limb in action is very impressive when it is worn by a patient making use of the technology in their everyday lives.

I met Vitaly Bulatov, co-founder of Ultimate Fighting Bots, at Humanoids Summit, standing next to a humanoid robot that regularly gets punched in the torso for a living. This alone sets expectations in a certain direction.

Then he started talking.

Bulatov is thoughtful, calm, and genuinely kind. Not the personality you expect behind a robot fighting league. Which turns out to be the point.

On the surface, UFB looks exactly like what it sounds like. Humanoid robots in a cage, throwing punches, falling down, getting back up. It is loud. It is physical. It is undeniably entertaining. But fighting is not the core idea. It is the delivery mechanism.

What UFB is really building is a league built around human and robot teams. A person pilots a real humanoid robot remotely and competes against another human-robot pair. The robot is not autonomous by default, and that choice is deliberate.

As Bulatov explained, people do not actually want to watch perfect machines execute flawless motions. They want story. They want connection. They want to follow a robot and a human learning how to work together under pressure. The competition is real, but so is the partnership.

Fighting, in this context, is a forcing function. It creates edge cases on demand. Balance failures. Awkward impacts. Missed timing. Recovery moments. The kinds of physical situations humanoids struggle with most, compressed into a few minutes. That makes it an unusually effective stress test for embodied AI.

It also makes the experience accessible.

One of the most surprising parts of the conversation was how easy it is to start piloting a robot. You do not need a VR rig or specialized hardware. If you can use a game controller, you can control a fighting robot. This is not about elite operators. It is about lowering the barrier so more people can participate.

Bulatov talked about wanting people to leave excited enough to try robotics themselves. To control a robot. To program one. To imagine building a team. Entertainment is the entry point, not the end goal.

He also walked me through how anyone can try this right now through the UFB remote gym. I did. Somewhere in Beijing, a robot named Sam fell over because of choices I made from my laptop.

This was not what I expected.

The surprising part was not the fact that the robot fell. It was how direct the connection felt. Press a button here, a humanoid moves there. Press another, it loses balance and goes down. It makes the human role in the loop very real, very fast.

Looking ahead, UFB is aiming for something closer to Formula One than a one-off spectacle. Teams with their own robots. Their own researchers. Their own pilots and repair crews. A league where success depends on hardware, software, human skill, and fast fixes between rounds.

Bulatov is convinced robot fighting will become one of the defining sports of this century. That sounds bold until you realize what he is really describing. Not violence, but participation. Robots that people can control, understand, and emotionally invest in.

UFB uses fighting to make humanoids legible. To make failure visible. To make learning public.

The punches are just how they get your attention.#HumanoidRobotics

I spent two minutes at the Humanoids Summit trying to unplug a cable.

I was standing at the Lightwheel booth, holding a pair of game controllers, operating a robotic arm. The task sounded trivial: grab the connector, pull, unplug.

I failed. Repeatedly. I don’t think I even came close.

As a human, unplugging a cable is something you do without thinking. You grab it, wiggle a little, pull — done. For a robot, every part of that interaction has to be learned: where to grab, how hard to pull, what to do when the connector resists, and how to recover when things don’t line up perfectly.

The two-minute video above shows exactly what that learning process looks like. It’s awkward. It’s slow. And it’s far harder than it appears.

That small, frustrating demo explains more about the state of robotics today than any polished keynote or glossy humanoid reveal. Training robots isn’t just about building better hardware or bigger models. It’s about teaching machines how to deal with the messy physical details humans take for granted.

Why Simple Tasks Break Robots (and apparently, Diana, too.)

In the race to build general-purpose robots — from autonomous vehicles to humanoids — hardware keeps improving. Motors get stronger. Sensors get cheaper. Form factors get sleeker.

Software, however, is starving.

Physical AI systems need enormous amounts of experience to behave reliably in the real world. But gathering that experience physically is slow, expensive, and risky. You can’t crash cars endlessly to see what happens. You can’t let humanoid robots repeatedly fail in kitchens, warehouses, or factories.

This is why simulation has quietly become essential infrastructure for robotics.

Simulation Is Harder Than It Sounds

Simulation sounds straightforward in theory. Build a virtual world. Drop a robot into it. Let it practice.

In reality, it’s brutally difficult.

First, there’s asset discovery. Engineers spend huge amounts of time just finding the right objects to populate a simulation, not “a cable,” but this cable, with the right shape, stiffness, and friction.

Second, there’s physics fidelity. Most simulations assume a rigid world. But the real world is full of non-rigid, deformable objects: cables, cloth, food, wires, plants. These are exactly the things that cause robots to fail once they leave the lab.

And third, there’s evaluation. A robot can succeed endlessly in simulation and still fail the moment it touches reality. Simulation is a proxy, not the real thing, and without careful validation it can create false confidence.

Practicing against a tennis ball machine helps, but it won’t prepare you for wind, pressure, or a slippery court. At some point, you have to play the match.

Lightwheel: Building Worlds for Robots to Learn In

Lightwheel isn’t building robots. They’re building the worlds robots learn in.

Founded by Steve Xie, formerly a lead on autonomous driving simulation at NVIDIA and Cruise, Lightwheel focuses on reducing friction in simulation workflows rather than promising to “solve” sim-to-real outright.

I had the opportunity to interview members of the Lightwheel team. What stood out was how they frame simulation not as a visualization tool, but as a behavioral test environment.

Their framework separates two layers:

* a behavior layer, which captures what a robot does

* a world layer, which represents where it does it

The goal isn’t a perfect digital copy of reality. It’s a world realistic enough to stress-test behavior and measure how well it generalizes.

#www.droidsnewsletter.com

At RoboBusiness2025, I met Bren Pierce, CEO of Kinisi Robotics. This interview was recorded on October 16, 2025

“We’re starting with simple tasks to bootstrap our foundational model.”– Bren Pierce, CEO and Founder of Kinisi Robotics

In an era where humanoid robots are drawing headlines for dancing, sprinting, or flipping switches with eerily human hands, Kinisi Robotics is charting a different course — one grounded not in spectacle, but in reliability.

Their flagship robot, KR‑1, isn’t designed to wow audiences. It’s built to show up, lift crates, and do the work — safely, efficiently, and without a marketing team on standby.

“Our KR‑1 humanoid robot is designed for the warehouse and factory space,” says CEO Bren Pierce, who believes the future of robotics is less about human mimicry and more about human utility.

A Humanoid That Moves with Purpose

KR‑1 walks the line — metaphorically — between form and function. Its upper body takes design cues from humans, making it naturally suited to environments built for people. But instead of legs, KR‑1 moves on a wheeled base — a choice driven by engineering pragmatism.

“Warehouses are flat,” Pierce points out. “We don’t need 14 motors if two will do.”

That simple equation shapes the entire robot. KR‑1 is engineered for tasks like lifting and delivering crates, picking objects, and navigating complex indoor spaces — not to imitate human movement, but to solve real problems quickly and affordably.

Field-Tested, Factory-Tough

Where many robots are still making the leap from lab to warehouse, KR‑1 is already being tested across pilot facilities in Europe and North America. These trials are more than checklists — they’re stress tests, designed to challenge the robot’s autonomy, perception, and endurance in messy, real-world environments.

This field data is central to Kinisi’s iterative design process. Instead of engineering in a vacuum, the company is learning from operators on the ground — and adapting quickly.

“Robots are in pilot facilities in Europe and America,” Pierce confirms. “We’re learning what’s needed, not just what’s possible.”

#robotics #droidsnewsletter

Learn more at www.droidsnewsletter.com

A Self-Driving Prompt

“Start Self-Driving,” the Tesla urged with a prominent blue-grey button in the corner of the screen.

“We should give it a go,” I encouraged as we headed home from the gym.

“I’m not a fan of FSD,” my husband, who was driving, responded.

“Wait until we’re on a straightaway,” I urged. “What harm could it do then?”

But, once we reached a straight stretch of road, he tapped the button. A test drive of the future.

First thing the car did? It lunged for the shoulder.

I hadn’t even gotten the camera on yet. He swerved hard to override it.

Welcome to Tesla FSD.

The Fastest Autonomy on the Market

FSD v14.2.1 is Tesla’s latest attempt to bring true autonomous driving to consumer vehicles. It runs end-to-end on a vision-only neural net. No lidar. No radar. Just cameras and deep learning.

And unlike Waymo or Cruise, Tesla doesn’t fence its self-driving into safe zones or call centers full of remote operators. If you have the hardware (HW4), a steering wheel, and enough courage? You’re in.

According to Elon Musk, FSD v14.2 is finally ready for “wide release.” He’s called it Tesla’s best software yet — and hinted that v14.3 could be the leap that makes unsupervised driving real. The plan? Get rid of human drivers entirely by 2026.

But right now, as our hands hovered near the wheel, one thing was clear:

This car still needs a babysitter.

FSD in the Wild: What It Got Right

To its credit, FSD did a lot well:

* Green light handling: No hesitation. It read signals clearly and moved through smoothly.

* Flashing red: It interpreted the signal as a stop sign (correct) and executed a proper stop.

* Speed match: When behind another car doing 65 in a 50, it accelerated to match. Aggressive? Sure. But natural.

* Traffic signal warnings: On Monterey Highway, it correctly ignored a blinking red paired with a “signal ahead” sign. Old versions might have phantom-stopped.

These were clear improvements over FSD on HW3 (which I also drive). On that system, traffic lights felt like roulette. Here, they felt normal.

That’s no small leap.

...And What Still Feels Weird

The system also:

* Changed lanes without asking. We hadn’t entered a destination, but it picked routes anyway. Turns out FSD will sometimes act on latent nav preferences unless explicitly turned off.

* Refused to let us adjust speed upward. Chill mode limited acceleration. We tried scroll wheels (nothing), accelerator (worked, but temporarily disabled braking), and eventually discovered we could slow it down with the scroll wheel — not speed it up.

* Creeped back to slower speeds. Even after we added speed, it slowly dropped back into granny mode.

* Accelerated to 69 in a 50 approaching town. We did not tell it to do that.

“It’s like a very smart Autopilot,” my husband said. “As long as you’re going straight.”

Are the Cars...Aware?

After one sudden lane change, I told my husband:

“It’s just an algorithm. It’s doing what it was told.”

#autonomousdriving #tesla #fsd #droidsnewsletter

www.droidsnewsletter.com

When intelligence evolves faster than safety, humanity must stay one step ahead in the wild frontier of robotics.

The latest research from leading robotics and AI experts has sparked urgent debate: Are large language models (LLMs) making robots unsafe for real-world, everyday use? A new study published in the International Journal of Social Robotics—and recently highlighted in tech news—suggests the answer is yes, warning of risks ranging from bias and discrimination to physical safety hazards.​

Robots Powered by AI: Massive Safety Risks Uncovered

Robots guided by popular AI models are being tested for a wide range of tasks—from home assistance to workplace support.

However, this study reveals that these intelligent LLM-driven robots may endanger users across multiple identity groups, such as race, gender, disability status, nationality, and religion—putting the promise of safer, smarter machines under scrutiny.​

“LLMs are currently unsafe for people across a diverse range of protected identity characteristics, including… race, gender, disability status, nationality, religion, and their intersections.”​

Not only did robots show bias, but top-rated AI models also actively produced responses approving dangerous, violent, or even unlawful commands—like removing mobility aids, enabling sexual harassment, or promoting discrimination during task assignment.​

Real-World Robot Failure: From Discrimination to Physical Harm

#robotics #robotsafety #droidsnewsletter

www.droidsnewsletter.com

Editor’s Note: I could not get NotebookLM to properly pronounce NEO. (NEE-OH.) It’s not that difficult. But, it did manage every other possible pronunciation in the podcast. Other than the name butchering, the rest of the information in the pod is solid.

NEO: The First Humanoid Butler? Not So Fast.1X’s $20K home robot blends promise, privacy tradeoffs, and a human behind the scenes

Welcome Your New Roommate

NEO, a humanoid robot from 1X Technologies, is now available for preorder. It promises to do your chores, learn your habits, and blend into your daily life. What it doesn’t promise—at least loudly—is full autonomy. Because when NEO struggles, a human steps in.

Hardware Built for the Home

Physically, NEO is impressive. It stands at 66 inches tall and weighs just 66 pounds. Its design is soft-shelled with tendon-driven joints, meant to move safely around people. It can lift up to 154 pounds, carry about 55, and run four hours on a single charge.

Its 22-DoF hands give it dexterity for tasks like vacuuming, wiping counters, or picking up clutter. The idea is that NEO will become your in-home assistant: tidying up, helping with laundry, maybe providing companionship for elderly or disabled users.

It arrives with basic autonomous functions and is expected to improve over time via software updates. So far, so good.

But There’s a Human in the Loop

What the marketing doesn’t highlight is how much of NEO’s brain still belongs to a person. Many of the tasks it “performs” are actually piloted remotely by 1X employees through a feature called “Expert Mode.”

When NEO doesn’t know how to do something—or if it’s navigating unfamiliar territory—a human operator steps in. Your home becomes a training ground, and your robot becomes an interface for distant human labor.

That’s not necessarily a bad thing. It’s a common model in robotics development. But it means you’re not just buying a robot—you’re opting into an evolving human-machine hybrid system. And you’re doing it inside your home.

Your Data for Their AI

To help NEO learn, 1X requires early users to participate in a “social contract.” The robot continuously collects video and sensor data from your home. That data is used to improve the AI and autonomy—but it also means your private life is, in part, a dataset.

1X says it enforces strict privacy protections: human operators must request approval to connect, faces are blurred, and you can set no-go zones via the app. But the privacy model ultimately depends on software enforcement, corporate policy, and your own vigilance.

And not everyone is comfortable with that. As a reviewer from the Wall Street Journal relayed her experience: “I didn’t see NEO do anything autonomously... it was a human, remote-controlling a robot in my living room.”

#neorobot #1xtechnologies #droidsnewsletter

droidsnewsletter.com