Humanoid robots have reached a new level of dexterity thanks to ASAP (Aligning Simulation and Real-World Physics), an innovative learning framework developed by Carnegie Mellon University and NVIDIA . This system improves the ability of humanoids to execute agile, coordinated full-body movements, overcoming the challenges of dynamic mismatch between simulation and reality.
Unitree G1 humanoid mimics difficult movements with precision
The Unitree G1 has proven capable of executing highly dynamic and challenging moves. It can replicate Cristiano Ronaldo’s mid-air spin with precision, mimicking his iconic “ Siu ” celebration. It has also managed to perform Kobe Bryant’s signature fadeaway , demonstrating advanced coordination in jumping and shooting. Additionally, it has successfully executed LeBron James’ “ Silencer ,” while balancing on one leg, as shown in the video below.
Beyond its sporting movements, the robot has managed to make forward and lateral jumps of more than a metre, improving its ability to react and move, as well as maintaining balance on uneven surfaces, adjusting in real time to changes in the terrain.
Stages of the ASAP learning framework
ASAP consists of four key steps:
- Pre-training : Reoriented human motion data is used to pre-train tracking policies in simulation.
- Delta Action Model : Real-world data is collected to train a model that compensates for the dynamic mismatch between simulation and reality.
- Fine tuning : The delta action model is integrated into the simulator to refine movement policies.
- Real implementation : The adjusted policy is deployed in the real world without the need for the delta action model.
Robots are becoming more agile
Training humanoid robots has faced significant obstacles due to differences between simulation and the real world. Traditional methods such as System Identification (SysID) and Domain Randomization (DR) often result in conservative policies that sacrifice agility. ASAP introduces a two-stage approach that dynamically adjusts motion policies to improve coordination and motion realism.
ASAP has been tested in a variety of environments, including IsaacGym , IsaacSim , Genesis , and the Unitree G1 humanoid robot , as seen in the video. The results show a substantial improvement in motion tracking accuracy, reducing error compared to traditional approaches. This allows humanoids to more faithfully replicate complex movements, such as mid-air turns and balancing on one foot.
Follow us on social media and don’t miss any of our posts!
YouTube LinkedIn Facebook Instagram X ( Twitter ) TikTok
Source and photos: Carnegie Mellon University and NVIDIA
