ASAP - A framework for aligning simulations with real physics, developed in collaboration between Carnegie Mellon University and NVIDIA.
ASAP (Aligning Simulation and Real Physics) is a two-stage framework jointly developed by Carnegie Mellon University and NVIDIA to address the mismatch between simulated and real-world dynamics in humanoid robots. ASAP is based on...
What is ASAP?
ASAP (Aligning Simulation and Real Physics) is a two-stage framework jointly developed by Carnegie Mellon University and NVIDIA to address the mismatch between simulated and real-world dynamics in humanoid robots. In the pre-training phase, ASAP learns motion tracking strategies in a simulated environment, generating target actions using human motion data. In the post-training phase, ASAP collects real-world motion data, trains a Delta motion model to compensate for the dynamic differences between simulation and reality, and fine-tunes the strategy to adapt it to the real physical environment. The framework significantly improves the agility and full-body coordination of humanoid robots in complex movements, reduces motion tracking errors, and enables efficient skill transfer from simulation to reality, providing a new technological path for developing more expressive and agile humanoid robots.
ASAP's main functions
- Transfer of Agility Full-Body SkillsThe goal is to efficiently transfer motion control strategies trained in a simulated environment to real robots to perform complex full-body movements, such as jumping, balancing, and rapid turning.
- Dynamic compensation and adaptationBased on the learned Delta motion model, it automatically compensates for the dynamic differences between simulation and reality, including hardware characteristics, friction, inertial parameters, etc., so that the robot can exhibit similar performance in the real environment as in the simulation.
- Improve motion tracking accuracyIt significantly reduces motion tracking errors when robots perform tasks, improves the accuracy and stability of actions, and maintains good performance in complex dynamic tasks.
- Enhanced generalization abilityThe framework is suitable for specific movements used during pre-training and can effectively adapt to unseen movements (i.e., "out-of-distribution"), enhancing the robot's generalization ability in the real world.
ASAP's technical principles
- Pre-training phase:
- Data generationThe process involves extracting motion data from human motion videos, re-localizing it onto a robot model, and generating a motion dataset for training.
- Motion tracking strategy trainingIn a simulated environment, a motion tracking policy is trained using reinforcement learning (RL) to enable the robot to track predefined motion trajectories. The policy is trained based on temporal phase variables to adapt to motion sequences of varying lengths.
- Post-training phase:
- Real-world data collectionThe pre-trained strategy is deployed onto a real robot to collect motion data during task execution, including information such as joint position, velocity, and acceleration.
- Delta Action Model TrainingBased on collected real-world data, a Delta action model is trained. The model learns how to compensate for dynamic mismatches by minimizing the difference between simulated and real states. Specifically, the Delta action model outputs a "delta action," which is added to the original action to adjust the dynamic characteristics of the simulated environment.
- Strategy fine-tuningThe trained Delta motion model is integrated into the simulator to fine-tune the pre-trained motion tracking strategy. Further training in the corrected simulation environment allows the strategy to better adapt to real-world physics, improving its performance on real robots.
ASAP project address
- Project official website:https://agile.human2humanoid.com/
- GitHub repository:https://github.com/LeCAR-Lab/ASAP
- arXiv technical paper:https://arxiv.org/pdf/2502.01143
ASAP application scenarios
- Sports and PerformanceRobots can simulate athletes' movements and be used in sports training, motion analysis, and performances to enhance visual effects.
- Disaster reliefThe robot performs agile movements, quickly reaches dangerous areas to carry out rescue missions, and improves rescue efficiency.
- Industrial manufacturingRobots can adapt to complex industrial tasks, improve production efficiency and flexibility, and reduce errors.
- Home servicesRobots can perform daily household chores, assist the elderly or disabled, and provide convenience and safety support.
- Virtual Reality and GamesGenerate realistic virtual character movements to enhance game immersion and development efficiency.