NVIDIA and Carnegie Mellon University launched a framework specifically designed for humanoid robots: ASAP
It can help robots accurately perform these actions in the real world after learning movements in a simulated environment, making highly flexible movements that were previously difficult to achieve possible!
The framework uses a two-stage process that allows robots to be pre-trained in a simulated environment, then applies these trained strategies to the real world, and achieves more flexible whole-body movements through further adjustments.
Bionic robots have unparalleled versatility to perform human-like whole-body skills. However, due to the dynamic mismatch between simulation and the real world, achieving agile and coordinated whole-body movements remains a major challenge. Existing methods, such as System Identification (SYSID) and Domain Randomization Method (DR) methods, often rely on labor-intensive parameter adjustments or policies that sacrifice agility that lead to overly conservative. In this article, we propose as soon as possible (Aligned Simulation and Real Physics), a two-stage framework designed to address dynamic mismatches and achieve agile humanoid full-body skills.
In the first phase, we used relocated human motion data to train a motion tracking strategy. In the second phase, we deploy policies in the real world and collect real-world data to train the Delta action model to compensate for dynamic mismatches. The pre-trained strategy is then fine-tuned with the Delta Action model’s ASAP and integrated into the simulator to effectively align with real-world dynamics. We evaluated ISAACGYM to Isaacsim, Isaacgym to Genesis and Isaacgym, and the real-world G1 Humoleoid Huoid Robot as soon as possible among three transfer scenarios. Compared to Sysid, DR and Delta Dynamics learning benchmarks, our method significantly improves agility and whole-body coordination across various dynamic movements, thereby reducing tracking errors.
Highly agile movements that were previously difficult to achieve were achieved as quickly as possible, demonstrating the potential of delta movement learning in bridging simulations and real-world dynamics. These results suggest that promising SIMs can be developed for more expressive and agile humanoid creatures into real directions.
There are four steps within the ASAP framework:
Motion tracking pre-training and actual trajectory collection:
Through human video repositioning of humanoid movements, we pre-trained multiple motion tracking strategies to extrapolate trajectories in the real world.
Irrigation action model training
Based on real-world launch data, we train the delta action model by minimizing the difference between the simulation state s_t and the real-world state s^r_t;
policy fine-tuning
We froze the delta motion model, incorporated it into the simulator to align with real-world physics, and then fine-tuned the pre-trained motion tracking strategy;
Real-world deployment
Finally, we directly deploy fine-tuning strategies in the real world without the Delta Action Model.
Address:https://agile.human2humanoid.com/
Thesis:https://arxiv.org/pdf/2502.01143
Oil tubing: