RoboHub🤖

World’s largest AI × robotics channel. Direct access to 200+ robotics companies. Original, tech-grounded reporting. 500K+ followers. 📬 [email protected]

[RoboHub🤖 @XRoboHub]

So glad we got to visit so many amazing robotics companies together.

Hope you come back every year. This industry moves fast and never disappoints.

There are so many great robotics companies here, and everyone is welcome to come see them.

[The Humanoid Hub @TheHumanoidHub]

That's a wrap on an amazing 17-day journey across China.

I didn't do this alone. My friend @dolylupec came along for the whole ride. She's got a background in AI and robotics investing, and I'm grateful she chronicled the tours on her account and pitched in on my channel.

@XRoboHub is one of the most resourceful people I know. He understands the humanoid space well and helped us schedule most of the important visits, then came along too.

And a shout-out to @Robo_Tuo, who encouraged me to do this trip and supported us all along.

It was my first time in China, and the trip was packed so tight with tours, we didn’t get to do much touristy stuff. Can't complain. It was a hell of a ride.

4 cities on the mainland. 18 humanoid makers. Warm, welcoming people everywhere. We had the privilege of chatting with founders, senior engineers, researchers, and marketing and strategy leads, a lot of whom said yes on very short notice. Grateful for that.

The energy is extraordinary. It's the trifecta: engineering talent, a deep supply chain, and breakneck adoption of new tech. There's intense competition, but underneath it, a real spirit of sharing and collaboration.

I've got a series of tour videos lined up, dropping over the coming days.

And something tells me this won't be my last trip. Robotics and embodied AI are just getting started. The next ten years are going to reshape how we think about physical work.

We ended on the Great Wall, a 2,000 year old marvel of engineering, after two weeks staring at the next one.

The robots are coming. I just went to meet them first.

[RoboHub🤖 @XRoboHub]

“From now on, anyone could become a roboticist.”

At CoRL, @DrJimFan laid out what that could look like: robots that are easier to teach through demonstrations and easier for AI agents to direct.

NVIDIA’s GR00T 2 handles the physical work. The team showed how skills demonstrated with human motion-capture gloves or handheld grippers can transfer to robots without collecting robot demonstrations for those specific tasks.

An AI agent can then oversee the job, check progress and respond to mistakes. In one assembly demo, it called a recovery routine to fix a misplaced part, then handed control back to GR00T 2 to finish.

Another robot worked through a lab procedure and cleaned up afterward. That last step matters when a researcher would otherwise have to reset the workspace before every run.

NVIDIA is also releasing data, software and hardware designs for other teams to build on.

GR00T 2 weights and X-Adapt training recipes are planned for December.

[RoboHub🤖 @XRoboHub]

@amanvaibhav Yep, the AI gets the spotlight, but good control is what makes it all work.

[RoboHub🤖 @XRoboHub]

If humanoids are going to work like us, they’ll need precise, stable control over their whole bodies.

Archon Robotics ’ Anchor 1 is a whole-body controller that turns movement targets from a human operator or robot AI into coordinated motion across the body.

It follows five targets: the pelvis, both wrists and both ankles. The controller coordinates the remaining joints and uses feedback to correct errors as the robot moves.

The team also built a benchmark to measure where the robot moves, how accurately it follows its targets, and whether it maintains that accuracy throughout a motion.

Across 72 real-robot motions, Anchor 1’s average global wrist position error was 75.69 mm, compared with 105.74 mm for MimicLite.

In a teleoperated box-kicking test, the team collected 43 successful demonstrations per hour with Anchor 1, compared with 9 using SONIC.

Archon also tested it with a fine-tuned GR00T N1.7 and plans to release the weights, deployment code and teleoperation tools by the end of 2026.

[RoboHub🤖 @XRoboHub]

@JonahVideos Yeah, that's a big step toward getting humanoids ready for real work!

[RoboHub🤖 @XRoboHub]

Most first-person video benchmarks for robot learning are way too coarse. A model can recognize “washing dishes” and still struggle to follow what each hand is doing.

Perceptron’s new EgoHome benchmark covers 137 first-person household clips with four annotation types: hand locations, task timelines, action labels and descriptions for each hand. It includes 773 annotated segments shorter than half a second.

GPT-6 Astra leads on hand-action labeling in Perceptron’s tests, but still scores just 0.498 out of 1.

Perceptron Egocentric roughly matches Astra on subtask labeling at about $15 versus $76 per hour of video.

[Perceptron AI @perceptroninc]

Introducing EgoHome: a new open benchmark for dense, sub-second egocentric understanding.

Human-annotated household-chore video with 4 annotation types on every clip: hand boxes, subtasks, hand actions, hand captions.

It features over 750 segments under 500ms.

Data, metrics, and judge prompts are all open. 🧵

[RoboHub🤖 @XRoboHub]

Robbyant just open-sourced a tactile model. They’re building a robot brain one sense at a time.

LingBot-Touch is @robbyant_brain’s JEPA-based tactile foundation model. The Ant Group company pretrained it on nearly 5,000 hours of recordings from tactile gloves and robot hands.

It turns signals from magnetic, optical and piezoresistive sensors into shared features describing contact location and force. Robot policies are trained to use those features while the touch model stays frozen.

Across 26 real-robot tasks, all four tested policies achieved their best average results with LingBot-Touch on both hands.

In the team’s evaluation, LingBot-VLA scored 48.3% without touch, 56.5% with raw tactile input and 73.1% with LingBot-Touch.

[Robbyant @robbyant_brain]

What changes when robots can feel—not just see? 🤖✋
Meet LingBot-Touch, our open-source tactile foundation model towards dexterous manipulation, pretrained on nearly 5,000 hours of real tactile data.
Through unified heterogeneous tactile modeling, LingBot-Touch supports transfer across sensors, dexterous hands, tasks, and action models.
Watch it in action. 👇 #LingBot #EmbodiedAI #Robotics

[RoboHub🤖 @XRoboHub]

@chris_j_paxton Nice, looks really cool!

[RoboHub🤖 @XRoboHub]

Source & more info 👇
https://t.co/BCVdcAuJ3H

[RoboHub🤖 @XRoboHub]

This little robot is already running errands. Give it a pair of hands and it’d be perfect.

PrimeBOT ’s T1 is a transformable robot that switches between wheeled humanoid and quadruped forms. In this demo, it takes the elevator downstairs, tosses out the trash and brings takeout back home.

The team says one guided walkthrough taught it the routine. It then completed the 21-step trip in 12 minutes and 32 seconds without remote control.

People still have to hang the bags on its arms. This is a research demo built on T1, not a standard feature of the retail robot.

In related research with Peking University, PrimeBOT’s One Scan is Enough uses an iPhone scan to build a virtual environment for training navigation, without collecting robot demonstrations in that space.

Across 200 real-world navigation trials in five environments, the team reports 85% success, versus 49% for the same model adapted with teleoperated demonstrations.

PrimeBOT has now produced 10,000 robots across its Q1 and T1 models, up from 5,000 just 14 days earlier. Looks like these robots aren’t just for fun—they can get real work done, too.

[RoboHub🤖 @XRoboHub]

This viral real-life Transformer could be hitting the market soon. 🤖

Rumor has it PrimeBOT’s T1 is already ramping up production and could launch within the next few months—priced lower than you’d expect.

What price would make you actually buy one?

[RoboHub🤖 @XRoboHub]

Source & more info 👇
https://t.co/McTjvHVkmo

[RoboHub🤖 @XRoboHub]

Apple’s robotics team is back with another new project. This time, they’re putting a general-purpose AI model in control of a robot arm.

It’s called AMP1, an agentic manipulation policy. Instead of relying on pre-programmed skills, the AI writes its own code to operate the robot and adjusts its actions based on what it sees.

The same framework works across single-arm, dual-arm and mobile manipulation, without task-specific demonstrations or fine-tuning.

On the RoboLab-120 simulation benchmark, AMP1 scored 85.6% with Gemini 4 Argon and 82.8% with Claude Opus 5.5. NVIDIA’s Astranomy, using GPT-6 Astra, scored 74.8%.

In real-world tests, AMP1 with Claude Opus 5.5 completed 104 of 108 trials across 12 tasks on three Franka setups. FLUX 3 Action completed 49, and π0.5 completed 30.

There’s an important difference in that comparison: AMP1 could recover from robot controller errors, while the baselines stopped. Without recovery, AMP1 completed 94 trials.

Speed is still a challenge. Successful AMP1 runs took 280 seconds on average, including 211 seconds for reasoning and related processing.

Apple has released the code for AMP1 and its simulation experiments.

[RoboHub🤖 @XRoboHub]

Apple’s robotics team seems to be preparing humanoids for the home, too.

In their demo, this Unitree G1 crouches under a table and turns sideways to get through a half-open door. It can also push a box, follow a person and sit down when instructed.

Behind it is HALO, an Apple research system that uses vision and language to generate whole-body motion. The team created 300 hours of training data across 1,000 simulated homes.

Across 360 single-task trials in six unseen real-world environments, none of which were pre-scanned, the team reports a 91% success rate. In a separate multi-task test in one real environment, it achieved 70% across 30 trials.

The team has also released code for generating training data, training the motion models and deploying them on a G1.

[RoboHub🤖 @XRoboHub]

@NVIDIA ’s new SimReady Central Library puts 3,500+ simulation assets on Hugging Face. @LightwheelAI contributed to the SimReady schema behind it and is a core contributor to the physics side of the library.

The company calibrated and validated assets for factories, homes and hospitals, and reviewed most of the collection before launch. Its assets include validated mass and inertia, collision geometry, and joint limits for articulated objects.

The engineering behind that starts with real objects. Lightwheel’s SimFoundry platform connects physical measurement, physics solving and asset generation: measure friction, stiffness and contact behavior, check the simulation against physical experiments, then produce reusable OpenUSD assets.

Its work on Newton’s Technical Steering Committee also includes SimReady schema development. The requirements are informed by the work of measuring, validating and producing the assets. For robotics teams, shared requirements can reduce the repeated work of configuring basic physics for each new training scene.

One concrete example is cable manipulation: Lightwheel uses real measurements and verification to tune assets for workflows where the robot’s motion and the cable’s deformation have to be modeled together.

[RoboHub🤖 @XRoboHub]

A company linked to Mandopop star Jay Chou teamed up with Unitree to create this adorable robot dog. It looks like a designer toy.

Called UNI-DOGGIE, it’s a companion robot dog from Star Plus Legend, with expressive eyes, floppy ears and a little tail.

It responds to touch and voice, follows its owner, and can dance, shake hands and give high-fives. There’s also a camera in its nose for photos and video.

Unveiled at the China International Import Expo, it costs 43,900 yuan (about $6,600) and is limited to 2,000 units worldwide.

The company was co-founded by Chou’s mother and manager and develops his official CHOUCHOU character. That connection, combined with the designer-toy appeal, could help turn some of his fans into robot buyers.

[RoboHub🤖 @XRoboHub]

Source & more info 👇
https://t.co/yzXyuhiNu6

[RoboHub🤖 @XRoboHub]

This is a social humanoid that can wave at you, shake your hand, and give you a high-five. It’s pretty cool to watch it interact with people.

Social Humanoid is a research project from UC Berkeley, Sony and Meta that trains a Unitree G1 to respond to a person’s movements in real time.

The team adapted 260 hours of human-to-human motion data for the robot, preserving interactions like handshakes and hugs despite the differences between human and robot bodies.

In real-world tests with four people, the G1 succeeded in all 30 trials in each of five interaction categories. High-fives succeeded in 27 out of 30 trials.

[RoboHub🤖 @XRoboHub]

@jianglong_ye Awesome project! Great job!

[RoboHub🤖 @XRoboHub]

The OpenRoboKit team has open-sourced a powerful set of tools for robotics developers.

RoboKit handles motion optimization and simulation. RetargetingKit maps human motion to robot hands, arms, and humanoids.

With GraspKit, developers can generate and test grasps in simulation, train a policy on the resulting data, and deploy it on real hardware.

There’s also RoboSkills, which gives AI coding agents tools and workflows for robotics tasks, including running simulation experiments and building interactive simulations from human demonstrations.

[RoboHub🤖 @XRoboHub]

Here’s their robot at work in lululemon’s Wuhan distribution center.

[RoboHub🤖 @XRoboHub]

The internet sees a robot having an emotional breakdown. 😂

Robotics people see a much bigger challenge: real-world manipulation.

This is X Square Robot's WALL-B embodied AI model powering autonomous parcel handling with a dual-arm robot.

Parcel sorting might look straightforward, but it requires robots to deal with different objects, unpredictable orientations, and continuous physical interactions.

The bigger question for Physical AI isn't whether a robot can perform a task once. It's whether foundation models can enable robots to handle diverse tasks reliably and generalize across real-world environments.

That's what makes systems like WALL-B interesting to watch.

We need more real-world robotics, not just polished demos.

[谢小总 @xie_Dl]

这工作连机器人都干得有情绪了…

[RoboHub🤖 @XRoboHub]

Source & more info 👇
https://t.co/1qbaHTSlZ8

[Jianglong Ye✈️CoRL2026 @jianglong_ye]

Excited to share OpenRoboKit, our series of open-source robotics projects! 🤖

- RoboKit for fast, scalable optimization & simulation https://t.co/9KosQmzZP3
Our RoboKit paper has been accepted for a spotlight at #CORL2026
- RetargetingKit for motion transfer & teleoperation https://t.co/fniCD3Uhms
- GraspKit for end-to-end grasping https://t.co/Tikg9Fz94U

Website: https://t.co/fzDLfxoPh5

Check the video (sound on 🎵) for a 16-hour grasping time-lapse in beautiful San Diego, plus a quick look at dexterous retargeting and robot simulation.