Meta The company announced Thursday the launch of a new product called Meta Motivo of artificial intelligence models.The model is designed to control the movements of humanoid digital intelligences to enhance theMetaverseExperience.
Meta also unveiled AI tools such as LCM, a large-scale conceptual model, and Video Seal, a video watermarking tool, and reiterated its commitment to continued investment in AI, AR, and metaverse technologies.
Note: Meta Motivo is a behavior-based base model trained in the Mujoco simulator using a subset of the AMASS motion capture dataset and 30 million online interaction samples pre-trained by a novel unsupervised reinforcement learning algorithm to control the motion of complex virtual humanoid intelligences.
Meta Motivo is trained using a novel FB-CPR algorithm that leverages unlabeled action datasets to bootstrap unsupervised reinforcement learning to learn human-like behaviors while preserving zero-sample inference capabilities.
Although the model was not explicitly trained for any particular task, performance on tasks such as movement trajectory tracking (e.g. cartwheels), pose arrival (e.g. arabesque dancing), and reward optimization (e.g. running) were all improved during pre-training, demonstrating behaviors that are closer to those of humans.
The key technological innovation of the algorithm lies in learning a representation that embeds states, actions, and rewards into the same latent space. As a result, Meta Motivo is able to solve a variety of whole-body control tasks, including motion tracking, target pose arrival, and reward optimization, without any additional training or planning.
refer to