UC Berkeley and Stanford unveil framework that teaches humanoids to run and perform spinning kicks from human motion data

Researchers at UC Berkeley and Stanford University developed a new AI framework called BeyondMimic to expand the range of movements humanoid robots can perform, according to an article published Sept. 10. The framework trains controllers using a large collection of human motion demonstrations, teaching robots agile movements such as running and spinning kicks.

As the researchers describe it, the framework aims to reduce movement-specific retuning, addressing the cost of training, which can be a bottleneck in humanoid control development. If it can handle movements such as running and spinning kicks, the focus will be not only on impressive demonstrations but also on how to transfer learning for a wider range of body movements to real robots.

The focus is not on humanoid hardware itself but on how control systems are trained. According to TechXplore’s account, BeyondMimic incorporates human motion data and combines it with reinforcement learning, with the aim of reducing movement-specific tuning and task-specific retraining. In other words, this appears to be an update to the learning infrastructure for expanding the repertoire of movements that can be used in the real world, rather than a competition over robot performance. The examples of running and spinning kicks suggest the framework targets movements that require balance control and continuous changes in posture, rather than simple repetitive motions. As Nikkei XTECH reports on a Unitree G1 teardown exhibit and GMO reports results from the World Humanoid Robot Games in Beijing, humanoid displays, competitions and evaluations are drawing attention. This article, however, is about the learning methods behind them. The issue is how to build control systems that can withstand real-world use behind the flashy movements seen at competitions. The comparison is therefore less about differences between robot hardware and more about the quality of training data and the ability of control systems to generalize. Publicly available information does not specify which robot was used, the size of the human dataset or the success rate on real hardware, so the framework’s generalizability remains to be assessed. The next points to check are the benchmarks, target robots and real-world test conditions presented in the paper.

Publicly available information does not indicate any direct connection to Japanese companies or robots. However, if data-driven control training becomes important in developing real-world humanoids, Japanese players may also face questions about how to build up capabilities in robot control, motion-data collection and real-world testing.