ASAP: Researchers have humanoid robots perform moves by famous athletes
A new framework better aligns simulations and reality in robot training. A Unitree G1 can thus imitate the moves of well-known athletes.
The Unitree G1 robot imitates Cristiano Ronaldo's "Siu" air jump.
(Image: LeCAR Lab / CMU (Screenshot))
A team of scientists from Carnegie Mellon University (CMU) and Nvidia has developed the AI framework Aligning Simulation and Reals Physics (ASAP) to perform iconic moves of famous athletes such as Cristiano Ronaldo and LeBron James with a humanoid robot. The researchers want to improve and exploit the agile full-body capabilities of humanoid robots.
Dynamic and coordinated full-body movements are no simple matter for humanoid robots. In most cases, agile movements are limited to locomotion or simple movement tasks. Recently, researchers have repeatedly tried to bring full-body agility to humanoid robots, but there are limitations due to the hardware and discrepancies between simulated and real physics.
To overcome these limitations, CMU scientists have developed the two-stage ASAP system for full-body agile humanoid capabilities with the help of Nvidia. The strategies are first trained in a simulation using human movement videos, as the researchers explain in the study "ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills", which has been published as a pre-print on Arxiv.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externes YouTube-Video (Google Ireland Limited) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (Google Ireland Limited) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
These strategies are then transferred to the humanoid robot. The robot is then trained with a motion tracking strategy that is compared with data from the real world. In order to compensate for these differences, the scientists train a delta action model. This makes it possible to fine-tune the strategies so that the robot can adapt to the real physical conditions. The researchers are thus closing the gap between simulated training and real execution. This makes the robots more agile.
More agile whole-body movements
The scientists used a Unitree G1 robot to test ASAP. They used videos to teach the robot typical movements of well-known athletes. Specifically, these were Cristiano Ronaldo's "Siu" loop with a twist in the air, the "Silencer", LeBron James' one-legged balancing act and Kobe Bryant's "Fadeway Jumshot". In addition, they trained him to perform various athletic movements, such as forward and sideways jumps of up to one meter.
The researchers found that by balancing simulated physics with real physics, the robot performed much more skillfully and was able to execute better whole-body movements.
Videos by heise
The researchers admit that the robot's movements still appear somewhat awkward, but this is due to the limitations of the hardware. The G1 robot simply has far fewer movable elements in its joints and its entire body than a human.
(olb)
