Physical AI: What's Working (and What Still Breaks)

Luma Event Link Physical AI is built in pieces: perception, mapping, simulation, policies, hardware. Every lab and startup holds one or two. Nobody holds them all. This night puts several of the pieces in one room, along with the people who built them. Robots are out on the floor when doors open. Then everyone sits down for talks and a panel about the parts that don't work yet: failure modes, sim-to-real gaps, the data that doesn't exist, and what happens where the pieces meet. Arrive at 5:30. Building security is located at the top of the Grand Staircase and can point you to the event. The robots are out from 5:30, and the program ends at 7:45. The lineup Niantic Spatial Punit Vats, Engineering  Real-world foundation models for physical AI: reconstruction accurate enough to train a robot in simulation. A five-minute capture of a room becomes a Gaussian splat carrying real lighting, texture, and clutter, with an aligned collision mesh, exported as a USDZ that loads straight into NVIDIA Isaac Sim. Hear Punit talk about how 3D Gaussian splats become the foundation for real-to-sim-to-real robotics workflows and policy training. Coco Robotics Rob Zehner, VP of Engineering Coco Robotics is transforming local commerce at the sidewalk level, offering customers and merchants a faster, lower-cost, and zero-emission delivery. Designed to respectfully blend into urban environments, our footpath delivery fleet operates at safe walking speeds, using multi-sensor spatial awareness paired with real-time safety oversight. By handling the last mile for restaurants, grocers, and neighborhood shops, Coco helps local businesses expand their reach and boost bottom lines while making city streets safer, quieter, and less crowded. At the August 25 event, we’ll be showcasing our hardware and software evolution, from Coco 0 robots hand-built by our founders in their college apartment to our newest Coco 2 vehicle designed for high-volume mass production and deployment in cities around the world. RLWRLD Carl Choi, President, US RLWRLD is a Physical AI company tackling the hardest challenge for the last mile of industrial automation: Dexterity. We build robotics foundation models that enable robots to perform precise, contact-rich tasks in environments designed for human hands. Our dexterity-first foundation model, RLDX-1, combines vision, motion, memory, tactile feedback, and force signals to help robots see, feel, remember, and adapt. The model is purpose-built for real-world deployments, engineered from a rigorous analysis of 5,000+ tasks across 200+ enterprises. Join RLWRLD alongside Niantic Spatial and Rally to hear candid insights from our ongoing proof-of-concepts & deployments, and experience a live demonstration of how RLDX-1 is helping close the dexterity gap in real-world industrial operations. How the night runs • 5:30 Doors and arrival. The robots are out at demo stations in the Spatial Station, with the people who built them standing next to them. Light charcuterie and sparkling water. • 6:00 Talks in the Forest. Each company takes ten minutes on what they built and where it still breaks. • 6:35 Panel and audience Q&A moderated by Lee Flannery, founder of Rally SF. • 7:20 Back to the robots, with open time at the demo stations.\ • 7:45 Event close. The Forest at Niantic Spatial. 1 Ferry Building, Suite 200, San Francisco. Getting in Niantic Spatial is inside the Ferry Building, which runs its own security, so everyone has to be on the approved guest list. Register with your full name as it appears on your photo ID, and bring that ID. Arrive at 5:30: Niantic staff will meet you at the bottom of the grand staircase and bring you up to Suite 200. The robots are out from 5:30, and the program ends at 7:45. Registration The room holds about 75, so registration is approved by hand. Please share a sentence about what you're building or why you're curious during registration. The full program is recorded and published on Rally SF's YouTube channel.…