Unitree G1 AI Robot Surgery Is Making EVERYBODY Panic ($41000 HUMANOID)

But here there’s a twist, because this process of transferring a policy from a simulated environment to a real humanoid robot involves overcoming the sim-to-real gap, as simulations only approximate the complex dynamics of real electromechanical systems. So to address this, a technique called domain randomization was used in simulation, where physical properties like friction or robot variations are randomized, training the policy to adapt to a wide range of conditions. This allowed the policy to generalize directly to physical robots without further adjustments. Additionally, a high frequency torque feedback control system was implemented on the robot to correct errors in actuator performance, ensuring the policy remained effective despite real-world variables like surface friction or external forces.

And this approach proved very successful, where a single reinforcement learning neural network enabled the consistent human-like walking across 10 robots without customization. And this scalability required no extra engineering effort, suggesting the potential for deploying thousands of Figure robots commercially in the future. And as a result of all of this, in just a few hours, Figure was able to run thousands of virtual humanoids through a series of scenarios inside a GPU-accelerated physics simulation to collect years worth of data. And all of that data was transferred directly from simulation to the real world, which is something that’s called zero-shot learning.

Tell me in the comments whether or not you trust this robot inside of your home, as well as whether or not you trust this next robot inside of your body, as number five is Sergi, the humanoid robot. This is a deployment of Chinese company Unitree’s G1 humanoid robot performing seven different medical procedures, thanks to researchers at ARC Lab in UC San Diego. Using a sophisticated teleoperation system, the G1 robot was outfitted with Inspire Gen 4 hands to mimic human movements with uncanny accuracy, controlled remotely via high-fidelity pose tracking, custom grasping configurations, as well as an impedance controller designed to handle delicate medical tools safely.

And the results were revealing, in physical examinations like oscillation where doctors listened to internal body sounds, the robot repositioned a digital stethoscope every 6.35 seconds, aligning it with anatomical landmarks to capture heart and lung sounds. However, its own vibrations muddled the signal, hinting at the need for finer tuning. Meanwhile, in obstetrics, the robot tackled Leopold maneuvers, a standard technique to assess fetal position during pregnancy, and while it successfully pressed and palpitated a maternal simulator, its rigid hands still struggled to replicate the nuanced touch of a seasoned clinician, and emergency scenarios offered even stronger results.

For bag-valve mass ventilation, which is a critical step to oxygenated patients in respiratory distress, the robot outperformed human consistency, whereas solo, it delivered breaths every 6 seconds with a tidal volume of 471 ml, hitting the target range 86.7% of the time. With both hands squeezing the bag and a human securing the mass, it averaged 533 ml, achieving 93.3% accuracy, which beat human performance, while maintaining a mass steel single-handedly still proved to be a bit more tricky. And the stakes were even higher with endotracheal intubation, where a tube is inserted into the trachea to secure an airway.

In this case, the robot’s strength wasn’t enough, needing a total of 44 newtons of force to open the patient’s mouth. Still, it placed the tube accurately once the airway was exposed, and in tracheostomy, a surgical airway creation, the robot cut and inserted a tube in under 20 seconds on average, though only 30% of incisions fully succeeded due to limited force and tactile feedback. And in precision needle tasks, the robot hit target vessels 45% of the time, with a 70% success rate when adjustments were allowed, which is performance that outpaces untrained medical students and trails closely behind expert clinicians.

But suturing exposed some further gaps, with the robot’s needle often veering off course into tissue, and the surgery robot’s teleoperators used HTC Vive trackers, multi-camera hand tracking, and foot pedals to command the robot’s arms in real time. Additionally, a virtual spring damper system sank the arms per cooperative tasks, while impedance control ensured gentle, precise contact. But the system still suggests that human touch is important, which is where this next Korean robot comes in with number 4, Egress from Robros in South Korea. This is the company’s newest 178-centimeter tall AI robot companion featuring 29 degrees of freedom throughout its body, with 6 degrees and 11 joints in its hands, allowing it to assist with everyday tasks like household chores or customer service, and the robot features a 48-volt dual battery system with 22 amp hours to carry the robot’s 106 kilogram weight, but its runtime and charge time both remain unspecified until now.

And the robot also features onboard AI hardware to power its natural language processing and facial recognition systems, allowing Egress to understand commands, recognize people, and adapt to different environments. But there’s still no word on pricing or release dates, and based on the robot’s 106 kilogram weight, it’s likely that the robot’s input cost will be significant, resulting in a retail cost of an excess of 100,000 US dollars. But it may have some competition from another company in South Korea, with number 3 from Rainbow Robotics. Its newly released RBY-1 is a dual-arm mobile manipulator with 7 degrees of freedom in each arm.

Its newly released RBY-1 is a dual-arm mobile manipulator with 7 degrees of freedom in each arm and a single leg with 6 degrees of freedom, all mounted on a humanoid-shaped, high-speed mobile platform. Designed for industrial applications, the RBY-1 overcomes the constraints of one-arm collaborative robots and fixed industrial models, enabling repetitive and precise tasks across various environments. And as for its performance, the RBY-1 stacks cups, pours water, and handles dynamic driving tasks, highlighting its real-world potential. But tell me how you think this robot will stack up against number 2, which is MagicBot from Magic Labs.

This company just released a new video showcasing MagicBot handling various consumer-facing tasks, including directing cars in an underground parking lot, serving as a sales assistant to a couple determining which car they want to buy, and giving a woman hair advice inside of a salon while it holds up a blow-dryer, as well as other natural language tasks. As for stats, MagicBot features 42 degrees of freedom with a payload of 40 kilograms in total. Its high-torque servo actuators, multi-dimensional pressure sensors, and 360-degree perception sensors offer submillimeter precision and obstacle avoidance, and they’re all powered by a proprietary navigation AI and the MagicData AI engine with a 5-hour battery life.

And this brings us to number 1, which is the S1 from AstroBot, and they just released a brand new video showcasing 5 times faster training using teleoperation for data collection. And this robot is seen handling a series of dexterous tasks while being controlled by its human operator. And the robot stands at a height of 173 centimeters and demonstrates extreme precision with a repeatability of just 0.03 millimeters, which bar surpasses human accuracy of about 1 millimeter, making it ideal for tasks requiring exceptional detail. But tell me in the comments which of these robots will rule them all.

Like and subscribe, and check out this video if you’d like to know more about the latest robots. [tr:trw].