AI Research Circle: Physical AI
85% of world GDP lives in the physical economy: moving goods, generating energy, building infrastructure, maintaining the systems that keep everything running. AI has transformed the digital 15%. Now it's moving into the rest.
Atharva Atre (Niantic Spatial) walks through the five technical layers making physical AI work:
Simulated worlds: training robots in synthetic environments before deploying them in reality (NVIDIA Cosmos)
Foundation models: one model that controls many different robots across many different tasks (Physical Intelligence pi0, Google Gemini Robotics)
Real-world reinforcement learning: grounding models in actual physical performance (Google DeepMind RT-2, Open X-Embodiment)
World models: learning the dynamics of how the physical world works (Yann LeCun's V-JEPA)
Spatial intelligence: mapping reality so machines can navigate it (Niantic's Large Geospatial Model)
We'll also cover who's building (NVIDIA, Physical Intelligence, WorldLabs, Skild AI, Niantic), where the capital is flowing, and what large-scale physical autonomy means for the workforce. B2B is already automating. B2C is where we'll start to feel it.
Format Conversational seminar. Each topic gets a simple explanation, a paper reference for anyone who wants to go deeper, and room for the group to riff. Memes included.
Suggested pre-reads
Presented by Rally SF, a community for AI builders in San Francisco. More events on Rally's Luma Calendar.