AI makes the most of robotics. A few months ago, we saw how robots properly iron their clothes, and now there are humanoids that can actually mimic the movements of your favorite soccer star. Carnegie Mellon and Nvidia researchers gathered to introduce the AI framework. This allows the robot to learn complex movements and simulations. The visual shows a robot mimicking an iconic festive film by professional athletes.
The ASAP framework stands for tuning simulations and real-world physics for learning agile humanoid whole-body skills. This is a two-stage framework designed to convert robot control policies from simulation environments to real applications, particularly for humanoid robots that perform agile whole body movements.
How does ASAP work?
Based on the research paper, the system works in two stages. Initial training in the simulation allows special neural networks to adapt movements suitable for real physics. The G1 robots from Chinese robotics company Unitree were tested by the team and featured complex movements, including regenerated animals from athletes such as Cristiano Ronaldo and Lebron James.
In stage 1 of ASAP, the framework trains the initial control policy in the simulation using motion data obtained from the video. For beginners, a robotics control policy is a set of rules or algorithms that tell the robot how to move and react to its environment. This kind of pre-training allows humanoid robots to easily mimic complex human movements.
In the second stage, when the robot is deployed into the real world, the ASAP framework collects data about the robot’s performance. The framework later trains a delta action model that compensates for the discrepancy between simulation and actual dynamics. Based on this, the model understands corrective actions and performs simulated actions based on actual responses, essentially tweaking the control policy.
The ASAP framework reportedly reduced motion errors by 53% when compared to some of the existing methods. This is seen as a major advance in streamlining virtual and physical training. However, hardware limitations remain a challenge as the two robots were damaged during testing due to the motor overheating while undergoing high-intensity movements.
This is a major leap in robotic motor functions. With training methods that advance with speed and efficiency, the robots can quickly play in the real field.
©IE Online Media Services Pvt Ltd
Enlarge
