HumanoidExo: Robots Learn Human Movements Faster Than Ever
Researchers have made significant strides in robotics, with a wearable suit called HumanoidExo enabling robots to learn human movements with remarkable efficiency. The device, developed by an unidentified research group, has facilitated rapid learning in humanoid robots, including walking and complex tasks like pick-and-place.
HumanoidExo maps seven human arm joints directly to a robot's configuration, allowing robots to mimic human movements accurately. The suit records a person's full-body motion, which is then converted into robot-ready actions using LiDAR, sensors, and AI models. This innovative approach has proven highly effective, with a Unitree G1 humanoid robot learning to walk after just five teleoperated demonstrations.
The robot achieved a 100% success rate in the locomotion portion and maintained balance throughout, thanks to the HumanoidExo-VLA, a Vision-Language-Action model that interprets tasks and maintains balance during movement. Training with exoskeleton data has also boosted success on complex tasks. For instance, it improved performance on a pick-and-place task from a mere 5% to around 80%.
The race to scale robot training is heating up, with tech giants like NVIDIA's Project GR00T and Google DeepMind's Gemini Robotics, along with startups such as Figure AI, competing to advance the field.
HumanoidExo's potential is undeniable, as it allows robots to learn new skills quickly and cost-effectively. By converting human motion into structured data, robots can gain new abilities, from walking to complex tasks, simply by observing people. As the technology continues to evolve, it promises to revolutionize the way we interact with and utilize robots in our daily lives.