Nvidia, XPENG, and CMU Bring Us Closer to a Future with Humanoid Helpers
New software platforms and rising competition is spurring the development of increasingly capable humanoid robots. | Credit: Pexels.
Key Takeaways
- In 2024, humanoid robots are increasingly capable.
- Chinese electric vehicle maker XPENG recently unveiled a new robot that could rival Tesla’s Optimus.
- New software platforms are accelerating the pace of innovation.
Humanoid robots have long captured the public imagination. But the practical challenges of recreating human-like movements and dexterity mean the concept has yet to break out of science fiction.
In the last month, however, advancements from Nvidia, XPENG, and Carnegie Mellon University (CMU) suggest the technology could be on the cusp of a major breakthrough.
2024: The Year of Humanoid Robots
After years of prototyping and a string of attention-grabbing but ultimately gimmick robots, in 2024, leading manufacturers like Boston Dynamics and Unitree showcased new models with impressive capabilities.
Perhaps the most notable trend of the year has been the breakout of Chinese companies, which positions the country as a heavyweight in the emerging humanoid robotics space.
Among these is the electric vehicle maker XPENG, which recently showcased a new bot the company said it has already started using in its factories.
Demonstrated during a company event on Oct. 31, XPENG’s “Iron” represents a significant leap forward compared to its previous prototypes and immediately drew comparison with Tesla’s Optimus. Discussing the project, CEO He Xiaopeng said advances in AI mean humanoid robots are developing faster and faster.
While more accessible AI has helped companies like XPENG close the gap with larger players like Tesla and Hyundai’s Boston Dynamics, new software platforms aimed at the robotics industry could further accelerate innovation.
For example, two new offerings from Nvidia and CMU could significantly lower the technical barriers for robot developers going forward.
Nvidia’s GR00T
Launched in March, Nvidia’s GR00T is a general-purpose foundation model for humanoid robots.
Reflecting the central role AI plays in the technology, GR00T is part of Nvidia’s broader push into the space, which has seen XPENG, Boston Dynamics, Unitree, and other leading players in the space utilize its robotics platforms.
On Nov. 6, the company introduced six new GR00T packages designed for challenges such as dexterous manipulation, perception for multimodal sensing, and mobility for locomotion and navigation.
Where GR00T is helping robot companies build the AI brains to power their machines, another tool developed by CMU researchers is bridging the gap between simulation-based training and real-world functionality.
CMU Researchers Introduce ManipGen
Models like GR00T must learn how to perform tasks successfully. However, training real-life robots is both time-consuming and expensive.
To overcome this challenge, developers use virtual simulations, essentially teaching the software to master tasks in a virtual environment before deploying it in a physical robot. Simulations, however, are inevitably imperfect models of the real world.
A team of CMU researchers recently developed ManipGen , designed to ease some of the friction between simulation and reality, for which they plan to release the code soon.
When teaching a robotic arm to perform 50 real-world manipulation tasks, the new software outperformed existing simulation training techniques by up to 76%.
Although it could theoretically be used for all kinds of robots, ManipGen may prove most valuable in the development of humanoid bots, which a heavy training burden has traditionally slowed.
China’s DeepRoute.ai Secures $100M To Deploy 200k Cars — A New Threat to Tesla?
