The Falcon system with a dynamic approach.
Researchers at Carnegie Mellon University developed technology called Falcon that is strengthening total body control in humanoid robots, this new approach promises to make robots much more resilient, stable and adaptable to the real world and by seeing the images of the robot pulling a cart, opening a heavy door or winning a tug-of-war competition, you can understand the pessimistic comments related to advances in technology and I read them.
With two AI “brains”
The Falcon is a reinforcement learning system with two artificial brains or rather two AI agents, one of them is responsible for stable locomotion, taking care of the legs, the other of the arms, it focuses on the precise manipulation of objects, this division of tasks allows the robot to walk firmly and at the same time execute complex actions that require strength, balance and coordination, such as pushing doors, pulling carts or lifting heavy objects.
The big difference is the way in which the Falcon trains the roads, it progressively increases the force applied to the robot's hands during learning, which avoids overload on the joints and guarantees precise control with a certain safety. During the tests the system was applied in two different humanoid robots Unitree G1 and Booster T1.
They tackled everyday tasks in unpredictable environments pulling carts with up to 30% of body weight, opening doors under pressure, and carrying payloads on the move, all while maintaining balance with 100% greater arm precision compared to previous methods.
Challenges to overcome.
Falcon also allows the same system to work on different robot models, without manual adjustments, it learns once and is applied to any structure which makes it easier to implement on a large scale, but not everything is perfect, Falcon now faces two important limitations, it concentrates only on the forces applied by the hands ignoring the rest of the body, that is why in tasks such as pushing with the shoulder or leaning against walls it still does not react as it should.
It has also not been trained to deal with torques such as turning handles or manipulating tools which requires another level of rotational control, despite these limitations Falcon marks an interesting advance in humanoid robotics the idea now is to integrate this type of learning with torque recognition and multi-touch reasoning, creating truly multifunctional robots to work in human environments from heavy industry to home care.
Comments