Meta AI has introduced Habitat 3, an innovative simulator designed to advance research on AI agents capable of cooperating with and aiding humans within virtual 3D settings. Developed by the Facebook AI Research (FAIR) team, this platform enables the creation of highly realistic human avatars and facilitates real-time interaction between humans and AI via a virtual interface.
Researchers highlight Habitat 3's exceptional ability to generate a wide range of lifelike virtual humans capable of navigating indoor environments and manipulating objects nearly as swiftly as non-human agents. In their accompanying research paper, FAIR describes Habitat 3.0 as a "simulation platform for investigating collaborative tasks involving humans and robots in domestic settings," emphasizing its contributions across three key dimensions.
Meta's overarching objective is to achieve "embodied AI," wherein technology actively participates in the physical world through wearable devices and robots. With the introduction of Habitat 3.0, Meta has established a platform designed for extensive training in human-robot interactions for everyday tasks, such as organizing a living space or preparing meals in the kitchen.
To assess these collaborative AI capabilities, the team devised two categories of interactive tasks. The first category involves social navigation, where a robot's objective is to locate and follow an avatar. The second category, social rearrangement, necessitates joint efforts from a robot and an avatar to tidy up a room. During their experiments, the researchers observed that virtual robots developed strategies and acquired behaviors to effectively collaborate with avatars, such as yielding space when navigating together. The social rearrangement experiments also demonstrated that these learned strategies could enhance work efficiency when dealing with previously unseen avatar partners.
Meta, in its announcement, acknowledged the substantial progress made toward realizing their vision of socially intelligent robots since they released Habitat 1.0 in 2019. However, they acknowledge that there is still important work to be done, and tools like Habitat 3 can stimulate increased engagement in this crucial research direction.
Looking ahead, Meta anticipates a focus on deploying the models trained in simulation into the physical world to assess their real-world performance.
In addition to this simulator, Meta has introduced the Habitat Synthetic Scenes Dataset (HSSD-200). This 3D dataset, created by artists, encompasses over 18,000 objects spanning 466 semantic categories across 211 scenes. It serves as a valuable resource for training navigation agents with unparalleled precision, offering a more realistic training environment for robots.
Meta has also unveiled HomeRobot, another platform designed to bridge the divide between simulation and the tangible world. HomeRobot enables robots to perform tasks in both simulated and real-world environments.
These advancements carry broad implications, particularly in the context of Meta's long-term objectives. The Metaverse, envisioned as a fusion of digitally enhanced physical reality and interactive digital spaces, stands to benefit from tools like Habitat 3.0, as they blur the line between the digital and physical realms, enabling more immersive and interactive experiences.
Furthermore, Meta's commitment to AI is evident through its substantial investments in the field. The company is a prominent contributor to platforms dedicated to natural language processing, underscoring its dedication to advancing the field. Its LlaMA LLM powers numerous open-source custom LLMs currently available, and it holds the record for the highest number of commits across open-source AI projects on Hugging Face, with Microsoft following closely behind.
Habitat 3 holds the promise of unlocking fresh research opportunities in the realm of assistive AI agents that can collaborate with humans in virtual environments, aligning with Meta's vision for collaborative AI in augmented and virtual reality. The ability to rapidly iterate and train policies in high-fidelity simulations before real-world deployment could accelerate the development of even more valuable AI capabilities.
