humanoid robots | Automation | AGI | ASI | Tracking the rise of autonomous systems and the future of intelligence Physical AI embodied
Etats-Unis
- Tweets 8,105
- Following 950
- Followers 18,086
- Likes 22,314
Pinned Tweet
For the first time I can actually picture humanoid robots being part of everyday life.
AGIBOT and Chimelong Group have brought 300+ robots to Chimelong’s resort in Zhuhai, China.
They’re taking on roles in performances, visitor guidance, companionship, education and hotel services.
The resort also received AGIBOT’s 20,000th humanoid robot, an A3 Ultra.
During the same stay, a guest could watch a robot perform, ask one for directions and meet others helping with hotel services.
I like seeing visitor guidance on that list. Finding your way around a resort is an ordinary part of a holiday.
Having a robot there to help gives people a straightforward reason to approach it and start a conversation.
For guests who have never been around humanoid robots, that could be their first encounter.
Asking where to go next.
300+ robots. One resort. Multiple real-world roles.
AGIBOT and Chimelong Group have officially launched a large-scale robot deployment at Chimelong in Zhuhai, China.
Across the resort, robots are being introduced for performances, AI education, visitor guidance, companionship and hotel services.
The launch also coincides with another milestone: AGIBOT’s 20,000th humanoid robot, an A3 Ultra, rolled off the production line and was delivered to Chimelong.
Built at scale. Now deployed at scale.
4
3
13
969
Humanoids are starting to assemble their own hands in the lab. We are cooking.
This five finger dexterous hand can handle grasping positioning and precision assembly with force feedback for more controlled contact with objects.
Seeing a humanoid work on components that belong to its own hardware is a pretty wild robotics moment.
For the first time I can actually picture humanoid robots being part of everyday life.
AGIBOT and Chimelong Group have brought 300+ robots to Chimelong’s resort in Zhuhai, China.
They’re taking on roles in performances, visitor guidance, companionship, education and hotel services.
The resort also received AGIBOT’s 20,000th humanoid robot, an A3 Ultra.
During the same stay, a guest could watch a robot perform, ask one for directions and meet others helping with hotel services.
I like seeing visitor guidance on that list. Finding your way around a resort is an ordinary part of a holiday.
Having a robot there to help gives people a straightforward reason to approach it and start a conversation.
For guests who have never been around humanoid robots, that could be their first encounter.
Asking where to go next.
1
2
7
361
A reinforcement learning agent learns from a number.
That number is the reward.
For a humanoid robot the hard part is deciding which behavior deserves a higher number.
Imagine training a humanoid to walk forward at 1 m/s.
A first reward could track commanded velocity:
r_velocity = −(v_target − v_robot)²
As the robot approaches the requested speed the error approaches zero.
Velocity alone can produce strange behavior.
The robot could move forward while leaning badly swinging its joints aggressively using large torques or making unstable contacts with the floor.
So humanoid locomotion usually combines several reward terms.
R = w₁r_velocity + w₂r_orientation + w₃r_height − w₄r_torque
Each weight controls how much one physical objective contributes.
Markov Decision Processes
Before PPO Actor-Critic or SAC, you need to understand the problem the policy is trying to solve.
That starts with a Markov Decision Process or MDP.
Imagine a simulated humanoid trying to stand.
At time step t, the simulator has a state s_t.
That state can contain things such as:
joint angles
joint velocities
pelvis position
body orientation
contact information
The policy receives information about the robot and chooses an action a_t.
For example the action could command joint targets for the hips knees and ankles.
The simulator then advances one step.
The robot moves.
Its feet may stay in contact with the floor.
Its torso may tilt
Its center of mass may shift.
The environment returns a new state s_{t+1} and a reward r_t.
Then the cycle repeats:
state → action → physics → next state → reward
An MDP assumes something very specific.
Once the current state and action are known, the probability of the next state does not need the full history of previous states.
Mathematically:
P(sₜ₊₁ | sₜ, aₜ)
For a robot, this assumption depends heavily on what information you include.
Suppose a humanoid is leaning forward.
Joint positions alone may show where the body is.
They do not tell you whether the torso is moving forward at 2°/s or 40°/s.
Those two situations can need very different actions.
That is one reason joint velocities angular velocities and contact states appear so often in robot control inputs.
A poor state description can make the learning problem much harder.
The same visible pose can lead to very different futures.
My favorite way to think about MDPs in robotics is to ignore the neural network for a moment and look at the physics loop.
The robot measures something.
It acts.
Physics responds.
The robot receives the result.
Every RL algorithm you learn later sits on top of that loop.
Tesla's current Optimus Whole Body Controls RL role includes designing reward functions action spaces observation spaces and curricula, then testing learned policies in simulation and on hardware. Those pieces all sit inside this interaction between policy and environment.
1
4
335
For a humanoid those reward terms could represent
• commanded forward velocity
• upright torso orientation
• target pelvis height
• joint torque cost
Isaac Lab exposes this structure through its Reward Manager which computes the total reward as a weighted sum of individual terms.
Its robotics reward functions include base orientation error base height error vertical velocity body acceleration and joint torque penalties.
A walking policy can therefore receive several objectives during the same time step.
Move at the requested speed.
Keep the body upright.
Avoid unnecessary vertical motion.
Control torque use.
Stay alive.
The weights determine how the policy trades these objectives against each other.
A large torque penalty can make very little movement produce a better return.
A strong survival reward combined with a weak walking reward can make standing still preferable to walking.
The policy follows the reward it receives.
1
1
87
Another quantity to understand is return.
Reward usually refers to the signal from one time step.
Return measures accumulated rewards over time.
Gₜ = rₜ + γrₜ₊₁ + γ²rₜ₊₂ + ...
γ is the discount factor.
A value close to 1 gives future rewards more weight.
This matters in robotics because actions often have delayed consequences.
A foot placement made now can determine whether a humanoid remains balanced several steps later.
OpenAI Spinning Up separates immediate reward from the return collected across a trajectory, which reinforcement learning methods try to maximize.
Tesla's Optimus Whole Body Controls RL role also lists reward functions action spaces observation spaces and curricula among the areas engineers work on.
The role covers learned policies for walking, balancing, disturbance recovery and object manipulation.
75
Markov Decision Processes
Before PPO Actor-Critic or SAC, you need to understand the problem the policy is trying to solve.
That starts with a Markov Decision Process or MDP.
Imagine a simulated humanoid trying to stand.
At time step t, the simulator has a state s_t.
That state can contain things such as:
joint angles
joint velocities
pelvis position
body orientation
contact information
The policy receives information about the robot and chooses an action a_t.
For example the action could command joint targets for the hips knees and ankles.
The simulator then advances one step.
The robot moves.
Its feet may stay in contact with the floor.
Its torso may tilt
Its center of mass may shift.
The environment returns a new state s_{t+1} and a reward r_t.
Then the cycle repeats:
state → action → physics → next state → reward
An MDP assumes something very specific.
Once the current state and action are known, the probability of the next state does not need the full history of previous states.
Mathematically:
P(sₜ₊₁ | sₜ, aₜ)
For a robot, this assumption depends heavily on what information you include.
Suppose a humanoid is leaning forward.
Joint positions alone may show where the body is.
They do not tell you whether the torso is moving forward at 2°/s or 40°/s.
Those two situations can need very different actions.
That is one reason joint velocities angular velocities and contact states appear so often in robot control inputs.
A poor state description can make the learning problem much harder.
The same visible pose can lead to very different futures.
My favorite way to think about MDPs in robotics is to ignore the neural network for a moment and look at the physics loop.
The robot measures something.
It acts.
Physics responds.
The robot receives the result.
Every RL algorithm you learn later sits on top of that loop.
Tesla's current Optimus Whole Body Controls RL role includes designing reward functions action spaces observation spaces and curricula, then testing learned policies in simulation and on hardware. Those pieces all sit inside this interaction between policy and environment.
This might be the most practical embodied AI model release I’ve seen this week.
Perceptron just released Mk1.5.
It can run across drones quadrupeds smart glasses and phones without retraining the model for each platform.
It takes text images video and audio.
It can track the same object through a video instead of treating every frame separately.
It can call tools like web search page reading and reverse image search. It can also spawn sub agents to work on tasks in parallel.
Perceptron reports a 36.1-point gain on MMSearch when tools are used.
And speed caught my attention.
Mk1.5 completes requests around 2–5× faster than Mk1, reaching up to 4.7× faster end-to-end in their tests.
For embodied agents that latency can matter when perception has to turn into an action fast.
Perceptron Mk1.5 is available now with a 32K multimodal context window.
#PhysicalAI #EmbodiedAI #Robotics #Perceptron
2
3
647
Practical exercise
Take a simulated humanoid standing task and write down five things:
State
pelvis orientation
joint positions
joint velocities
base angular velocity
foot contact states
Action
desired joint positions or joint torques
Transition
the physics simulator moves the robot forward one simulation step
Reward
a scalar produced after that transition
Episode termination
the robot falls or the episode reaches its time limit
Then remove joint velocity from your state representation and think about what information disappears.
That small exercise will make the Markov assumption much easier to understand before touching PPO.
Berkeley's RL material describes reinforcement learning using an MDP where actions produce successor states and rewards, while OpenAI's RL notes formalize trajectories through the transition distribution P(s_{t+1}|s_t,a_t).
Image idea:
A side-view humanoid inside a physics simulator with a circular five-step diagram:
State
joint angles + velocities + contacts
↓
Policy
↓
Action
joint targets or torques
↓
Physics simulator
↓
Next state + reward
Use arrows to show the loop repeating every control step.
Sources:
Tesla, Reinforcement Learning Engineer, Whole Body Controls, Optimus
"Tesla careers listing" (https://reference-url-citation.invalid/2)
UC Berkeley CS188, Reinforcement Learning
"Berkeley CS188 RL notes" (https://reference-url-citation.invalid/3)
OpenAI Spinning Up, Key Concepts in Reinforcement Learning
"OpenAI Spinning Up RL introduction" (https://reference-url-citation.invalid/4)
Progress:
Covered: Markov Decision Processes
Still remaining in Reinforcement Learning:
rewards
observations
action spaces
policy gradients
Actor-Critic
PPO
SAC
exploration
reward shaping
curriculum learning
parallel environments
Reinforcement Learning is still in progress.
1
127
This might be the most practical embodied AI model release I’ve seen this week.
Perceptron just released Mk1.5.
It can run across drones quadrupeds smart glasses and phones without retraining the model for each platform.
It takes text images video and audio.
It can track the same object through a video instead of treating every frame separately.
It can call tools like web search page reading and reverse image search. It can also spawn sub agents to work on tasks in parallel.
Perceptron reports a 36.1-point gain on MMSearch when tools are used.
And speed caught my attention.
Mk1.5 completes requests around 2–5× faster than Mk1, reaching up to 4.7× faster end-to-end in their tests.
For embodied agents that latency can matter when perception has to turn into an action fast.
Perceptron Mk1.5 is available now with a 32K multimodal context window.
#PhysicalAI #EmbodiedAI #Robotics #Perceptron
Today we're releasing Mk1.5: a new intelligence layer for embodied agents.
It flies drones, controls quadrupeds, powers smart glasses, tracks objects, searches the web, reasons visually, and dispatches its own sub-agents.
One model, no platform-specific retraining. 🧵
3
3
12
1,109
What caught my attention with Mk1.5 is how Perceptron is treating the model as the decision layer across different physical systems.
The same model can take video, audio and images track objects over time call external tools and then choose actions exposed by a drone quadruped or another device.
The next thing I’d watch closely is how this holds up in longer real-world tasks where latency recovery and repeated tool calls start to stack up.
1
1
149
France has a small but diverse humanoid robotics scene
• Wandercraft Calvin-40
Bipedal humanoid built for industrial work
• UMA Northstar
General-purpose humanoid under development in Paris
• Enchanted Tools Mirokaï
Mobile humanoid robots designed for service and assistance
• Pollen Robotics Reachy 2
Open-source humanoid platform for manipulation and Embodied AI
• InMoov
Open-source humanoid created by French designer Gaël Langevin
• NAO
Historic French humanoid developed by Aldebaran
• Pepper
Social humanoid originally developed by Aldebaran
• Romeo
Full-size humanoid research project from Aldebaran
France currently has fewer humanoid companies than China or the US, but its projects cover industrial work, research, healthcare, service robotics and Physical AI.
#HumanoidRobot #Robotics #PhysicalAI #EmbodiedAI #France
I started mapping out everything inside a humanoid hand, and the list got long very quickly.
You have the fingers and joints for movement.
Motors, gearboxes and tendons generate force. Tactile, force and position sensors track contact.
Then the control system has to coordinate all of it during a grasp.
And that is just to pick something up, hold it, adjust the grip and move it without dropping or crushing it.
Seeing all these parts together made me realize how much engineering sits behind one simple hand movement.
A humanoid hand is a small system with a lot going on inside.
1
1
10
886
A humanoid robot hand looks simple until you see what is packed inside.
Control boards, finger mechanisms, joints and tiny mechanical links all have to fit inside a space close to the size of a human hand.
Then every finger has to move together while the thumb follows its own path for grasping.
Watching the hand get rebuilt piece by piece made the engineering difficulty much easier to see.
Five fingers on the outside. A lot of coordination underneath.
Watching a humanoid control policy make the jump from simulation to physical hardware never gets old.
This is exactly what reinforcement learning and sim to real training looks like in practice.
The digital environment gives the robot a place to fail thousands of times while it masters balance and locomotion.
All those awkward falls and recoveries happen safely without putting any wear and tear on the expensive physical machine.
Having the simulated motion side by side with the real robot is the best part.
You can instantly see how the exact same movement logic translates once actual surface friction and physical actuators get involved.
1
10
884
Watching a humanoid control policy make the jump from simulation to physical hardware never gets old.
This is exactly what reinforcement learning and sim to real training looks like in practice.
The digital environment gives the robot a place to fail thousands of times while it masters balance and locomotion.
All those awkward falls and recoveries happen safely without putting any wear and tear on the expensive physical machine.
Having the simulated motion side by side with the real robot is the best part.
You can instantly see how the exact same movement logic translates once actual surface friction and physical actuators get involved.
2004 vs 2026. Same company, completely different robot.
Boston Dynamics went from BigDog stumbling across rough terrain to the electric Atlas built for industrial work.
22 years of robotics progress in two clips.
Watching them back to back makes the jump feel huge.
7
1,836
2004 vs 2026. Same company, completely different robot.
Boston Dynamics went from BigDog stumbling across rough terrain to the electric Atlas built for industrial work.
22 years of robotics progress in two clips.
Watching them back to back makes the jump feel huge.
I started mapping out everything inside a humanoid hand, and the list got long very quickly.
You have the fingers and joints for movement.
Motors, gearboxes and tendons generate force. Tactile, force and position sensors track contact.
Then the control system has to coordinate all of it during a grasp.
And that is just to pick something up, hold it, adjust the grip and move it without dropping or crushing it.
Seeing all these parts together made me realize how much engineering sits behind one simple hand movement.
A humanoid hand is a small system with a lot going on inside.
2
7
39
5,167
This might be one of the best projects ever released in human history.
Being able to carry a powerful multimodal intelligence inside a small box already sounds unreal.
Muse can understand images, language and instructions, plan tasks and work with robotic control systems.
Now imagine connecting that brain to a humanoid.
Cameras become its eyes.
Microphones become its ears.
Tactile sensors become its sense of touch.
The VLA controls the hands and arms.
Real time controllers handle balance and movement.
You could talk to the robot normally, give it a long task, and let the system break that task into physical actions.
Meta has already tested Muse as a high level robot brain and orchestrator for manipulation research.
A full humanoid running Muse has not been announced yet.
But seeing this intelligence move from a screen into a physical body would be something else.
The main downsides
Latency: a large multimodal model may react too slowly for situations where a humanoid needs millisecond-level control.
Power consumption: running a strong model locally can drain the robot’s battery much faster.
Heat: GPUs and onboard compute add heat inside a body already packed with motors, batteries, and electronics.
Compute size: some models may still be too large to run fully on the robot, which can force part of the system into the cloud.
Network dependence: cloud-based reasoning becomes a problem when the connection is weak or unavailable.
Reasoning errors: misunderstanding a scene or instruction is far more serious when the system controls a physical machine.
Motor control limits: Muse could plan tasks, but balance, torque control, hand motion, collision response, and fast reflexes still need dedicated real-time controllers.
Sensor failures: bad lighting, hidden objects, reflective surfaces, noise, or faulty sensors can lead to the wrong decision.
Safety: strong physical robots need hardware-level force limits, emergency stops, collision detection, and restricted motion zones.
Privacy: a humanoid with cameras, microphones, and memory could collect large amounts of personal data inside homes and workplaces.
Cost: stronger compute, cooling, batteries, and sensors all add to the hardware cost.
Real-world reliability: a system may work well in controlled tests and still fail when the floor changes, an object moves, or a person suddenly enters its path.
The hardest problem may be the gap between understanding a task and carrying it out safely with the body.
1
4
621
I started mapping out everything inside a humanoid hand, and the list got long very quickly.
You have the fingers and joints for movement.
Motors, gearboxes and tendons generate force. Tactile, force and position sensors track contact.
Then the control system has to coordinate all of it during a grasp.
And that is just to pick something up, hold it, adjust the grip and move it without dropping or crushing it.
Seeing all these parts together made me realize how much engineering sits behind one simple hand movement.
A humanoid hand is a small system with a lot going on inside.
Something changed fast with Unitree’s humanoids.
Back in June, you could spot the robots on stage almost immediately.
By September, they were dressed in the same uniforms as the human performers.
By late September, I genuinely couldn’t tell which dancers were the Unitree R1 and H1 while watching the crowd.
I only knew once the robots were revealed.
A few months ago, my eyes went straight to the robot because its movements gave it away.
Now there are moments where it blends into the choreography well enough that I stop looking for the machine and just watch the performance.
That gap between human and robot movement is getting harder to spot on stage.
3
6
5,156
Something changed fast with Unitree’s humanoids.
Back in June, you could spot the robots on stage almost immediately.
By September, they were dressed in the same uniforms as the human performers.
By late September, I genuinely couldn’t tell which dancers were the Unitree R1 and H1 while watching the crowd.
I only knew once the robots were revealed.
A few months ago, my eyes went straight to the robot because its movements gave it away.
Now there are moments where it blends into the choreography well enough that I stop looking for the machine and just watch the performance.
That gap between human and robot movement is getting harder to spot on stage.
TENNIIX and LimX Dynamics just unveiled ULTRA MAX an embodied AI tennis robot built on the TRON 2 wheeled-leg platform.
TENNIIX says it tracks ball trajectory, spin, landing point and player position, then adjusts the next shot based on the return.
It can feed from 1.2 to 1.6 meters and move around the court between shots.
They also put it on court with Wang Qiang at the Billie Jean King Cup Finals.
I’d like to see shot-placement accuracy and how many rallies it can complete before human intervention.
2
3
14
3,446
TENNIIX and LimX Dynamics just unveiled ULTRA MAX an embodied AI tennis robot built on the TRON 2 wheeled-leg platform.
TENNIIX says it tracks ball trajectory, spin, landing point and player position, then adjusts the next shot based on the return.
It can feed from 1.2 to 1.6 meters and move around the court between shots.
They also put it on court with Wang Qiang at the Billie Jean King Cup Finals.
I’d like to see shot-placement accuracy and how many rallies it can complete before human intervention.
Tesla’s Optimus Gen 3 is starting to show up inside the Tesla app.
The new renders show more covered hips and legs, with a much more finished design.
Some people already think this could be the next Optimus model.
Tesla is also moving on the manufacturing side. Fremont has been converted, supplier audits are happening in China, and work on the Texas plant structure is moving forward.
But production is still very limited. The head of sensors for Optimus also just left Tesla, while consumer sales are still being discussed around 2027.
What catches my attention is the distance between how finished these Gen 3 renders look and how few Optimus units are actually being produced today.
So where are you right now on Optimus?
Mostly Tesla hype, or a robot you expect to actually see working in factories and eventually homes?
1
4
1,791
Tesla’s Optimus Gen 3 is starting to show up inside the Tesla app.
The new renders show more covered hips and legs, with a much more finished design.
Some people already think this could be the next Optimus model.
Tesla is also moving on the manufacturing side. Fremont has been converted, supplier audits are happening in China, and work on the Texas plant structure is moving forward.
But production is still very limited. The head of sensors for Optimus also just left Tesla, while consumer sales are still being discussed around 2027.
What catches my attention is the distance between how finished these Gen 3 renders look and how few Optimus units are actually being produced today.
So where are you right now on Optimus?
Mostly Tesla hype, or a robot you expect to actually see working in factories and eventually homes?
GENISOM AI just landed a massive Series B funding round of several hundred million RMB led by Stone Venture in the UAE.
They have been quietly building a really solid ecosystem featuring quadruped robots their NE01 humanoid joint modules and the navigation software that powers them.
According to the team, they already pushed past 15,000 manufactured units by June putting these robots to work in seriously demanding environments like power plants petrochemical facilities and emergency response zones.
Since it carries a robust explosion proof rating (Ex d IIC T6 Gb) specifically for inspecting gas and petrochemical sites, the real proof will be in the data.
It would be incredibly useful to see them publish some concrete numbers on actual patrol durations fault rates and route completion percentages to see how it is holding up in the field.
7
15
91
12,498
GENISOM AI just landed a massive Series B funding round of several hundred million RMB led by Stone Venture in the UAE.
They have been quietly building a really solid ecosystem featuring quadruped robots their NE01 humanoid joint modules and the navigation software that powers them.
According to the team, they already pushed past 15,000 manufactured units by June putting these robots to work in seriously demanding environments like power plants petrochemical facilities and emergency response zones.
Since it carries a robust explosion proof rating (Ex d IIC T6 Gb) specifically for inspecting gas and petrochemical sites, the real proof will be in the data.
It would be incredibly useful to see them publish some concrete numbers on actual patrol durations fault rates and route completion percentages to see how it is holding up in the field.
Zhongyi Embodied raised nearly RMB 100M in an angel round led by the controlling shareholder of Andar Intelligent, with Shengang Hengtian and Beijing Jungeri participating.
The Shenzhen company doesn’t manufacture robot bodies.
It builds the software layer that connects and schedules mixed fleets of humanoids, quadrupeds, cleaning robots and delivery robots.
Its TianGongJi platform already brings together 40+ robot manufacturers.
That would give a much better view of how this type of robot OS performs in real buildings.
1
4
12,303
Zhongyi Embodied raised nearly RMB 100M in an angel round led by the controlling shareholder of Andar Intelligent, with Shengang Hengtian and Beijing Jungeri participating.
The Shenzhen company doesn’t manufacture robot bodies.
It builds the software layer that connects and schedules mixed fleets of humanoids, quadrupeds, cleaning robots and delivery robots.
Its TianGongJi platform already brings together 40+ robot manufacturers.
That would give a much better view of how this type of robot OS performs in real buildings.
A future where some people choose a robot over a human partner no longer feels impossible.
Look at what AheadForm is already building.
Its Origin F1 has 26 degrees of freedom across the head, eyes and face. It can recognize faces, listen, speak and respond with facial expressions.
Then there’s Elf Xuan, with more than 30 facial degrees of freedom and silicone skin that moves with the face.
Once a robot can look at you remember previous conversations notice how you speak and respond with small expressions the interaction starts feeling much closer to companionship than using an AI app on a phone.
Give this hardware another decade of progress in memory speech and behavior and some people may genuinely prefer an AI companion living beside them.
That sounds strange today.
It may feel completely normal to the next generation.
1
1,008