Amit Ezer (@AmitEzer) header
Amit Ezer (@AmitEzer) profile picture

Co-founder & former CEO of Upswift, an IoT & Robotics platform acquired by @JFrog Now building a startup to keep the next-gen of Physical AI fleets running.

Joined November 2017
Photo by Amit Ezer (@AmitEzer)Photo by Amit Ezer (@AmitEzer)Photo by Amit Ezer (@AmitEzer)Photo by Amit Ezer (@AmitEzer)Photo by Amit Ezer (@AmitEzer)Photo by Amit Ezer (@AmitEzer)Photo by Amit Ezer (@AmitEzer)Photo by Amit Ezer (@AmitEzer)Photo by Amit Ezer (@AmitEzer)
Speech to action is becoming real. Did someone say Physical AI is here? 👀
True drone autonomy on top of almost any other flight platform is now here Degla is already working with mission critical government agencies for search & rescue
1
4
460
I bet most software engineers think building a humanoid robotics startup is way out of their league. THINK AGAIN. You don't have to build the robot or train the AI model yourself anymore. Find a job that humans do today, buy a programmable robot, get access to a robotics foundation model, and start connecting the dots. Sounds too simple? It is. Making it work in the real world is still incredibly hard. But getting started has never been easier. And to make your life easier, I put together a list of robots you can buy, preorder, or build: 🇺🇸 NORI A3 — $1,688 (preorder) 🇨🇳 VLAI L1 — $3,974 🇨🇳 AgileX LIMO COBOT — $4,400 🇨🇳 Unitree R1 — $4,900 🇨🇳 Booster K1 — $5,999 🇺🇸 StarBot Spark — $7,999 🇺🇸 Weave Isaac 1 — $7,999 (preorder) 🇺🇸🇨🇦 MABEL — ~$8,000 (DIY) 🇨🇳 Elephant Robotics myArm M&C — $9,499 🇺🇸 StarBot Astra — $9,999 🇰🇷 ROBOTIS OP3 — ~$12,000 🇨🇳 Unitree G1 — $13,500 🇺🇸 OS3 HALE — $14,995 🇩🇪 NEURA 4NE1 Mini — €19,999 🇺🇸 RC Mobile ALOHA — $20,000 🇸🇬 Menlo Asimov 1 — $20,000 (kit) 🇨🇳 LimX Oli — ~$21,800 🇨🇳 AgiBot X2 Neo — $25,600 🇨🇳 Unitree H2 — $29,900 🇺🇸 Feather — $29,990 🇺🇸 ARX LIFT3 — $29,999 🇨🇳 AgileX COBOT MAGIC — $30,110 🇨🇳 Booster T2 — ~$33,000 🇺🇸 Trossen Rivet — $38,480 🇺🇸 Lightberry Lumi — $39,990 🇨🇳 Galaxea R1 Lite — $39,999 🇨🇳 Fourier N1 — ~$40,000 🇺🇸 Dexmate Vega U — $45,000 🇺🇸 Fauna Sprout — ~$50,000 🇺🇸 RC M1 — $50,000 🇰🇷 ROBOTIS AI Worker — $69,000 🇨🇳 Galaxea R1 Pro — $69,999 🇫🇷 Reachy 2 — ~$70,000 🇺🇸 Dexmate Vega — $83,000 Some are full humanoids, others are wheeled robots with arms. Availability, prices, and support for custom AI models vary. A few years ago, you'd have to build almost everything yourself. Now you can start with the hardware and intelligence others have already built. What a time to build!
8
7
3
128
6,326
From your Dyson vacuum to a robotic arm doing pick-and-pack, every dollar in hardware matters. Now add Physical AI, and that problem gets 10X harder. That's why this is VERY COOL. @physical_int π0.5 is still one of the most impressive open-source VLAs out there. And someone just built microVLA, a π0.5-style model that's ~286x smaller than the original! AND it still achieved 90% success on a robotics manipulation benchmark in simulation. BOM (Bill of Materials) is one of the biggest enemies of physical products. Having to put a $1,500 or even $3,500 AI compute system inside a product can completely change its economics. Now imagine running useful VLAs on $150–$500 compute modules. A whole new category of products could become AI-powered, Right? To be clear, this is a smaller, task-specific model trained from scratch, not a compressed version of the original π0.5. We're not there yet, but I believe we'll see huge progress on this front very soon
Introducing microVLA! A minimal pure pytorch implementation of pi0.5 VLA training and inference on a RTX 3050. Github link below. Please do check it out, and RT/QT for the reach:)
4
3
15
928
This is the kind of stuff that makes me want to jump out of bed every morning and work on robotics. Let me explain why this is so exciting. An MRI lets doctors see inside the human brain in incredible detail. But regular robots can't operate inside one because of its powerful magnet. These guys at 4D Surgical are building a robot that can actually work inside an MRI, using live brain imaging to guide treatments with incredible precision. Think about how many lives technology like this could improve or even save. I really hope I'll get to work with teams like this and help them bring their technology reliably into the real world. Watch this video and tell me the future isn't exciting
Bring a normal robot near an MRI and this happens That’s why the best way to see the brain and the most precise way to operate in it have never worked together We just raised $8M @4DSurgical to change that
1
1
8
960
I'm starting to think Tesla FSD's biggest competitor might be Optimus sitting in the driver's seat
2
1
111
We're witnessing the GPT moment for robotics, driven by a wave of new startups. AI-powered, off-the-shelf robots. Everywhere. Every industry. Don't let anyone tell you otherwise
Dad never taught you to change a tire? Say less. Rodeo Robotics. Robots for vehicle maintenance.
2
304
There are a few pretty unusual GTM decisions here that I kind of love, but also find hard to wrap my head around. The more I looked into it, the more questions I had. 1. It's an Android phone. But they're marketing it as an entirely new category of phone, going straight after iPhone users. Barely any mention of Android in the main pitch. Bold positioning. 2. Why sell an entire phone? Their biggest innovation is software. But instead of asking users to download an app, they're asking them to switch phones entirely. I understand that OS-level access might be necessary, but I wonder if an Android launcher could achieve at least some of what they're trying to do. For the iPhone people here, a launcher is basically an app that lets you replace Android's interface and customize parts of the experience without replacing the OS. Either way, that's a HUGE distribution decision. 3. The hardware is literally a Google Pixel 10 Pro/XL. Running their modified Android OS. They're selling it for $1,299, but offering it for free if you trade in your iPhone. With subscriptions planned from $20/month, I'm genuinely curious how the unit economics work. I can't decide if this is genius or a very expensive way to acquire users. Either way, it's one of the more interesting GTM bets I've seen from a startup lately. Congrats to the team, and good luck!
We switched 10 people from iPhone to Blue. Weeks later, none of them want their iPhone back. Introducing Blue: a phone you just talk to. Tell it what you need, and it gets done. It has all the context, can use every single app, and even triages your notifications. We built it for people who always have a little too much to do. The ones buried in messages, notifications, and to-dos. If that's you, get yours today. Backed by Y Combinator + 70 other VCs & angels. heyblue.com
2
5
420
Come on. A company that makes AI video generators is now building robot brains?! What is even happening in Physical AI 😂 And of course, they're taking the WAM approach I explained yesterday with NVIDIA. No rest in this field. Let me quickly explain why this is interesting. Most traditional VLAs learn to turn what a robot sees and what you ask it to do into physical actions. Runway takes a completely different starting point. They start with a massive video generation model trained to predict how the world changes over time. Then they add a smaller model, trained with robot data, that learns to generate movements alongside those imagined futures. Think about it. The robot is predicting what the next few moments will look like AND how to make them happen. And apparently, it's working. Their new Praxis-1 just beat every open-weight VLA on NVIDIA's RoboLab-120 simulation benchmark. Yesterday I was trying to figure out whether VLAs or WAMs would win. Today an AI video company enters the race and tops the open-weight leaderboard. I officially have no idea where to place my bet anymore 😅
Today, we’re making Praxis-1 available to everyone. Our state-of-the-art world action model turns video pretraining into control for real robots, and it ships with open weights under Apache 2.0, ready to run and fine-tune on your own hardware.
5
374
Every time you're tempted to buy a car that isn't a Tesla, watch this and remind yourself why there's really no other choice
66
So many big things are happening in Physical AI these days, and this is one of them, even if it's easy to miss. Let me explain what a WAM is and where it stands in the race to build the best robot brain :) Most robotics foundation models we've been hearing about are built around VLAs (Vision-Language-Action). The idea is simple. A robot sees a cup, gets an instruction like "put it on the table," and the model predicts the movements needed to do it. Basically: See → Understand → Act. But NVIDIA's new GR00T 2 is taking a different approach. It's called a World-Action Model (WAM). And here's the interesting difference. While a typical VLA learns to predict actions from what the robot sees, a WAM learns to predict TWO things together: How the robot should move. What the world will look like as it moves. Imagine picking up a cup. The model learns both the movements needed to pick it up AND how the cup, hand, and surrounding scene will change. Not two separate steps. Both predictions are learned together. Now, why does this matter? Because learning how the physical world changes opens up another powerful source of training data: human videos. And to be clear, VLAs are learning from human videos too. @physical_int and @GeneralistAI are already exploring this. What's interesting about WAMs is HOW they learn from that data. NVIDIA trained GR00T 2 on over 400,000 HOURS of human videos, alongside robot and handheld-gripper data. And the results? On 50 tasks seen in human or other robots' data, but never demonstrated on the target robot, GR00T 2 achieved 40% success without task-specific training. Give it just FIVE demonstrations per task and a 30-minute fine-tune? 95% success. That's pretty incredible. We're watching several different approaches to robot intelligence evolve at the same time
Introducing GR00T 2: our most advanced, open System One model for robotics. It runs entirely on your robot’s onboard compute, fully untethered, and you own every weight of it. GR00T 2 features: 1. Long-horizon tasks: dexterous manipulation and whole-body control chained over 20+ steps and 5+ minutes, such as performing chemistry lab experiments or assembling model airplanes end to end. 2. X-Adapt: adapting from easy data sources (human wearables and simulation) and deploying zero-shot to diverse real robots. 3. Unified Cognition: native integrations with frontier agents for advanced physical reasoning, robust error recovery, and safety monitoring. It's an open-source buffet! We are releasing not only the model weights and X-Adapt recipes (coming December), but also our agentic framework, simulation engine, annotation engine, hardware design, and even training data (100K hours, that's ~11 yrs of experience), all available today! Everything that moves will be autonomous. Anyone can become a roboticist. Lock in, build with us! Deep dive in thread:
5
365
Here we go! This is exciting. Google, Apptronik, and Mercedes-Benz showing us the first signs of Physical AI making its way into real-world production. Remember this post. A few months from now, I bet we'll start seeing hundreds of deployments like this. The glass ceiling is finally breaking.
Replying to @GoogleDeepMind
In a proof of concept with @Apptronik and @MercedesBenz Gemini Robotic 2.5 helped the Apollo 2 robot pick and organize car parts for assembly with 96% accuracy.
1
1
11
1,080
I don't know... I just burned $500 in Fable 5.2 tokens trying to get it to control my robot arm, and nothing seems to work well. Sorry, but robotics foundation models are in a league of their own. Tho I'd love to see both approaches competing
Replying to @claudeai
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
2
1
20
1,970
One thing I think people underestimate about Physical AI: updating an AI model isn't just a software release anymore. You're changing how a physical machine behaves in the real world.
1
1
157
There's a suspicious correlation between men suddenly trying to convince their wives to have a third child and the release of the 6-seat Tesla Model Y L. Elon is out there planning to launch data centers into space and somehow also fixing America's birth rate. Unbelievable.
115
I think we're about to see an explosion of billion-dollar Physical AI startups. And most won't build their own hardware or AI models. Hear me out for a second. What if you could buy existing robot hardware, take a robotics foundation model, post-train it for a specific task, and deploy it? Construction, agriculture, manufacturing, you name it. As hardware gets cheaper and foundation models get better, the barrier to building a robotics company is changing dramatically. My bet? We'll see tons of startups doing exactly this, and many will build incredible businesses. The real opportunity might not be building the next robot or foundation model, but figuring out what to do with them.
53
25
15
393
32,275
Physical AI research doesn't seem to slow down, and I keep finding myself explaining another interesting development. This time, it's something called SLA (Spatial-Language-Action). Researchers from HKUST took @physical_int 's π0.5 robotics model and made it significantly better at real-world manipulation tasks. So what's it all about? In simple words, today's robotics foundation models are pretty good at recognizing objects, but understanding their positions in 3D space is still a challenge. The researchers used a separate 3D vision model during training to help π0.5 learn better spatial representations. In one real-world experiment, a human deliberately moved a piece of fruit while the robot was reaching for it. The success rate jumped from 42% to 76%. And here's the interesting part: the extra 3D model was only needed during training. The robot doesn't need it when deployed. Pretty interesting direction for improving Physical AI models without making them bigger at inference time
16
868
Two years ago, this would have been an incredibly ambitious project, even for a well-funded robotics company. Today, we're seeing small startups take on some of the hardest real-world tasks, powered by robotics foundation models and increasingly capable hardware. The pace of progress in Physical AI is crazy
Introducing Theseus Robotics (YC F26). Our robots restocks supermarket shelves overnight and picks online orders by day, without changing anything in the store. Lights off at 9pm, shelves full by 9am. Next stop: San Francisco.
20
2,072
This is seriously impressive. Humanoids are coming to the real world much faster than most people expect
Meet AIR, Embodied AI’s new wheeled humanoid built for factory work. Its tendon-driven architecture keeps the motors, gearboxes and electronics in the torso, leaving the arms lightweight and compliant. It starts under human control, learns from demonstrations, then progressively gains autonomy. A practical approach to physical AI.
5
275
Here's a crazy problem Physical AI is creating: the computer can cost as much as the entire robot. I've seen edge devices in every shape and size, across different industries. For years, the computing hardware in many of them cost $50–$500. Now we're entering an era where running a robotics foundation model on-device can require something like @nvidia 's Jetson Thor, a $5,000 module. That's roughly the starting price of an entire Unitree humanoid. We're all excited about how intelligent these machines are becoming. But I think making that intelligence affordable at scale is going to be one of the most interesting challenges in robotics
1
77