Rashid Arif

Building remote teams since 2011. Now building Labelix: dedicated in-office teams powering AI and physical AI. Real teams, clean data, not crowds

[Rashid Arif @ratrashid]

One of the things I enjoy most about working in robotics is that it feels like everyone is still learning together. There are new demos, companies, and ideas almost every week. It's exciting to watch the industry evolve in real time.

I'm curious, what recent development in Physical AI has caught your attention?

We also shared a few thoughts about it in our latest blog: https://t.co/DR0pqSuIDz

[Rashid Arif @ratrashid]

@clavion95 The repetitive part of construction looks like a natural fit for robots. The harder question is how well they handle everything that isn't perfectly set up for them.

[Rashid Arif @ratrashid]

@NeuralPulsaulq The consistency is impressive. Reliable physical work at scale is where robots could really make a difference.

[Rashid Arif @ratrashid]

10,000+ autonomous delivery bots. At this scale, robotics starts looking less like a pilot and more like infrastructure.

[Ronald van Loon @Ronald_vanLoon]

Inside China's Massive Fleet of 10,000+ #Autonomous Delivery Bots
by @XueJia24682

#SupplyChain #Innovation #FutureTech #EV

[Piggle @pixu_alpha]

Unitree burned $4M to make their G1 humanoid beat pro drummers in real-time rhythm

The result is terrifyingly good, mastering micro-second timing required training neural motor control from scratch.

Instead of pre-programmed loops, the robot processes dynamic feedback on the fly to hit exact percussion velocity and timing.

Why this physical AI benchmark actually matters:

Sub-Millisecond Precision: Real musical rhythm demands instant trajectory adjustments standard actuators fail at.

Spatial Dexterity: Maps multi-surface impact points instantly while maintaining acoustic balance.

[The Humanoid Hub @TheHumanoidHub]

Top moment from last year's World Humanoid Robot Games.

The operator jogging alongside the robot picked the exact wrong moment to hand off the controller…

[Techman Robot @techmanrobot]

📍 Live from Booth #M620 on Day 1 of Automation Taipei! See AI move beyond concepts and power real-world applications. Visit us and experience smart manufacturing in action! #AutomationTaipei2026 #AI #TMAICOBOT #TechmanRobot

[Léo @LeoKharon]

NEW ROBOT SKILL: Your humanoid can now skateboard!

This project involves @jinrui82 and others.

Called HUSKY, it is a learned whole-body controller that makes a Unitree G1 humanoid (23 controllable DoF) ride a real skateboard, pushing forward with one foot and steering by leaning.

A few DEFINITIONS:
- Underactuated: the system has more ways to move than motors to command them. The robot can't directly control the board's motion and must steer it indirectly through its own body, the way you can't directly command a swing, only pump it.
- Non-holonomic: constrained in which directions it can move at any instant -> like a car, the skateboard can roll forward and turn but never slide sideways, so reaching a nearby point can require a long curved path.
- Adversarial Motion Priors (AMP): instead of copying a reference motion frame by frame, a discriminator network is trained to tell robot motion from human motion-capture, and the policy is rewarded for fooling it. The robot learns the style of a human push without being chained to one exact trajectory.
- Rake geometry: the skateboard truck's kingpin is mounted at an angle, which mechanically converts deck lean into wheel steering; that's why the policy never commands a steering angle, it just leans and the geometry does the rest.
- Plant: control-theory jargon for the physical system being controlled. "Each board is a new plant" means each deck's flex and damping must be measured and plugged in separately, because the controller is tuned to those numbers.
- PD controller (the 500 Hz layer): a simple low-level loop that turns the policy's 50 Hz joint targets into motor torques by correcting position and velocity errors ten times faster.

Instead of treating the deck as fixed ground, it models the humanoid and skateboard as one coupled, underactuated, non-holonomic system connected only through intermittent foot contact, where the truck steering angle is set by lean through fixed rake geometry (tan sigma = tan lambda sin gamma).

The policy has three parts:
1. An adversarial Motion Priors learn a human-like pushing stroke
2. A physics-guided law converts a desired heading into a closed-form lean reference
3. A trajectory-guided mechanism handles the switch between the one-foot push stance and the two-foot ride stance.

It runs at 50 Hz (500 Hz PD) trained in mjlab (MuJoCo physics with the IsaacLab API) and is deployed on the real G1 indoors and outdoors.

With two real decks (one stiffer, one more compliant), each needs its own identified stiffness and damping; cross-applying the wrong board's parameters causes failure.
The robot's own dynamics transfer fine under domain randomization, but the passive object's compliance does not, and because steering is lean-driven the whole controller is exquisitely sensitive to the tilt spring-damper, so each physical board is effectively a new plant.
-> That is a sharper sim-to-real statement than randomize-and-pray.

It steers with zero commanded steering angle, purely by leaning.
The policy never actuates the truck angle; it regulates lean toward a physics-derived reference, and rake geometry converts lean into steer.
In the ablation, removing that physics tilt reference does not kill success (96.72 vs 100 percent) but degrades heading precision (yaw error 0.233 vs 0.208).
So the physics prior buys precision of an emergent skill, not feasibility, the robot can flail into steering, but the closed-form reference makes it accurate.

Motion-tracking-based pushing collapses to 11.12 percent success against AMP's 100 percent; a fixed-gait push gets 82 percent but with 130 times the contact violation (0.130 vs 0.001).
On a moving, underactuated platform, rigidly tracking a reference push trajectory fights the board's dynamics, whereas the adversarial/style formulation lets the stroke adapt to the board state, which is what keeps it upright at all.

One caveat: no brakes!
Propulsion is push-only and speed is regulated during the push; stopping/deceleration is never modeled or discussed, presumably left to coasting and friction.
The camera field of view cannot see the board or the wheel-ground contact, so the policy rides essentially blind on proprioception, a real deployment gap.

[Humanoids daily @humanoidsdaily]

Day job vs. Weekend plans. 📦🥊

On the left: EngineAI’s T800 quietly handling payloads at a Luxshare-ICT factory.

On the right: The exact same bot brawling for its life at @REK in San Francisco over the weekend.

High-torque stress-testing meets industrial automation.

[Rashid Arif @ratrashid]

@CGTNOfficial 9.58 seconds is an ambitious target. 😂 Getting a humanoid to sprint that fast without losing stability is the real challenge.

[Jason Smith - 上官杰文 @ShangguanJiewen]

China's droid army has arrived.

Imagine tens of millions of homes, each with its own droid helper; that's China in 2030.

[Space and Technology @spaceandtech_]

Wandercraft launched Calvin-40 last year as a humanoid robot designed for industrial work.

It was developed in just 40 days using technology from the company’s medical exoskeletons.

The robot is designed to help with heavy lifting and repetitive tasks in factories.

[Rashid Arif @ratrashid]

@SciFry01 The force sensing is the part that gets me. Getting a robot to handle something as delicate and variable as meat is no small feat.

[Rashid Arif @ratrashid]

@jacksonhinkle This is where humanoid robotics starts getting real: actual logistics work, at scale.

[Rashid Arif @ratrashid]

@cbames @REK I think I’ll give this one plenty of personal space. 😂

[el.cine @EHuanglu]

the speed of AI robot in china rn

[Rashid Arif @ratrashid]

Robot training needs better data, and capturing human movement this way could be a big part of it.

[Space and Technology @spaceandtech_]

General Intelligence Labs has unveiled EGO1GS, an egocentric headset designed to capture high-quality data for training robots.

It features global-shutter stereo cameras, stereo audio, and a 400 Hz IMU for capturing human movements.

Its EgoHand AI model runs directly on the headset and detects the wearer’s hands in real time to help ensure higher-quality robot-training data.

[Rashid Arif @ratrashid]

Seeing harvest automation operate at scale is a big deal. Agriculture is one of those areas where robotics can have a very practical impact.

[Danny Bernstein 🍇🥬 @bernsteinres]

robotics in agriculture is becoming more real. for growers, harvest automation is their #1, #2, and #3 priority.

this is @BeagleTechAI operating in celery harvest at scale in California’s Central Coast.

[The Humanoid Hub @TheHumanoidHub]

A battalion of Booster T2s is getting ready for the World Humanoid Robot Games opening ceremony.

Now in its 2nd year, the Games run Aug 22-26 in Beijing at the "Ice Ribbon" Olympic oval.

It's the biggest yet: ~2,056 robots from 660+ teams across 16 countries, competing in 51 events.

Beyond track-and-field style sprints and jumps, this year adds scenario events like robot tug-of-war, weightlifting and pitch-pot. Part sporting event, part real-world capability benchmark for humanoids.

[Rashid Arif @ratrashid]

Brain-controlled robots are a wild direction for human-robot interaction. And under 200ms is seriously fast.

[Space and Technology @spaceandtech_]

China-based BrainCo has unveiled a system that lets people control robots using their brain signals.

A lightweight EEG headset detects brain activity, which AI decodes into commands for the robot.

The process takes under 200 milliseconds, and the platform can work with commercially available robots.

[Science Magazine @ScienceMagazine]

Humanoid robots can army crawl, breakdance, and do continuous backflips without task-specific training, using a new motion-imitation framework called ZEST.

Learn more in @SciRobotics: https://t.co/9nJSEWANag