AI is bringing the best out of robotics. While months earlier we saw how robots were ironing clothes neatly, now we have humanoids that can actually imitate the movements of your favourite soccer star. Researchers at Carnegie Mellon and Nvidia have come together and introduced an AI framework, ASAP which lets robots learn complex movements and simulations. The visuals show robots imitating some iconic celebration movies by professional athletes. ASAP framework stands for Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-body Skills. It is a two-stage framework that has been designed to transform robotic control policies from simulation environments to real-world applications, especially for humanoid robots which perform agile whole-body motions. How does ASAP work? Based on the research paper, the system works in two stages - initial training in simulation after which a specialised neural network adapts movements suited for real-world physics. Chinese robotics company Unitree's G1 robots were tested by the team and it showcased complex motions including recreating moves from athletes like Cristiano Ronaldo and LeBron James. In stage 1 of ASAP, the framework uses motion data taken from videos to train initial control policies in simulation. For the uninitiated, control policies in robotics are simply a set of rules or algorithms that tell a robot how to move around and react to its environment. This kind of pre-training lets the humanoid robots imitate complex human movements with ease. In the second stage, when the robot is deployed in the real world, the ASAP framework collects data on how the robot performs. The framework later trains a delta action model which makes up for the mismatch between simulation and real-world dynamics. Based on this the model understands corrective actions to carry out simulated actions based on real-world responses essentially fine-tuning its control policies. View this post on Instagram A post shared by Rowan Cheung (@rowancheung) Reportedly, the ASAP framework reduced errors in motion by 53 per cent when compared with some of the existing methods. This is seen as a major advancement in streamlining virtual and physical training. However, it seems hardware limitations remain a challenge, as two robots suffered damage during tests due to overheated motors while performing high-intensity movements. This is a big leap in robotic movement capabilities. With training methods advancing with speed and efficiency, robots may soon play in the actual field.
