A robot sees a glass and a towel. The difficult part is understanding every physical action between them. This Perceptron Egocentric demo captures a task that demands continuous coordination. ➝ One hand stabilizes the glass ➝ The second hand folds and reshapes the towel ➝ The grip changes as the towel enters the glass ➝ Contact continues while the object rotates ➝ Pressure shifts across the inner and outer surfaces ➝ Each action receives its own temporal boundary ➝ Both hands remain tracked through the full sequence The result is structured supervision built from a normal human task. This type of egocentric data can help manipulation models learn the exact movements behind household actions instead of relying on one broad video caption. Early access at #EmbodiedAI# #RobotLearning# #PhysicalAI# #Robotics#

Humanoid robots look simple from the outside. Inside, they are a dense stack of machining, bearings, electronics, sensors, cables and thermal constraints. This breakdown shows why humanoid robotics is hard before the robot even starts walking. • CNC precision machining Used for rigid structures where tolerance matters Shoulder housings Waist joints Hip joints Knee housings Lower leg frames Foot structures A small alignment error can affect balance, actuator wear and repeatability. • Bearings Used across almost every rotating joint Shoulders Waist Hips Thigh actuators Knees Force sensor modules Crossed roller bearings are critical because humanoid joints carry radial, axial and moment loads during motion. • Injection molding Used for shells, covers, gloves and outer body parts. These components reduce weight, protect internal hardware and give the robot a clean exterior. • PCB electronics Used for the sensing and control layer RGB camera Depth camera Encoder boards Mission computer Sensor PCB Pressure sensor board This is where perception, feedback and low-level control begin. • Cable harnesses The hidden failure point. Cables pass through moving joints. They bend thousands of times. They must survive vibration, heat and repeated walking cycles. Key components shown here: • LiDAR • RGB camera • Depth camera • Mission computer • Shoulder actuator • Encoder • Electromagnetic brake • Upper arm rotary actuator • Robotic hand • Waist joint • Hip joint • Thigh actuator • Knee joint • 6-axis force sensor • Lower leg structure • Foot pressure sensor • Cable harness AI gets the attention. But the physical stack decides whether the robot can work for hours without broken joints, overheated electronics, loose cables or damaged sensors. Physical AI needs a physical supply chain.

The Unitree G1 executed kicks forward jumps and squats after ASAP trained policies from retargeted human motion and corrected the sim to real gap with rollout data from the physical robot. The key detail is the hardware loop. The team measured how the real G1 responded trained a delta action model from that data and used it to reduce the mismatch between MuJoCo and the robot. This matters because humanoid control often breaks when the foot hits the floor. A jump or fast leg swing can expose errors in timing friction torque and balance. ASAP shows how learned motion can move from simulation onto real hardware with less collapse at contact.

A humanoid robot grabs the kendama and goes straight to work. Both arms lock in sync. In five days of Physical AI training it transformed. Now it is stable and fast like industrial machinery. It is closing in on the certified 135 reps per minute. One hand clamps the striped handle. The other drives the red ball through sharp controlled arcs. Blue lights stay fixed on its head while every motion stays tight and intentional. #Unitree# #humanoid# #robot# From: @yaman_xr