I’ve been watching Physical AI data scale from tens of thousands of hours to millions. But seeing that Maxinsights has collected 2 million hours of people doing chores in real homes genuinely surprised me. I haven’t seen many teams build a household dataset at this scale. Every home robot company wants robots to handle everyday chores, yet even clearing dishes or folding a pile of laundry can still be a challenge. Real-world training data has been scarce. Of those 2 million hours, 1.5 million are already ready for training. And that still won’t be enough. 5M, 10M, 50M hours…?humanoid robots have a huge appetite for real-world experience. We’re in the Physical AI gold rush....The race is on to see who can capture the physical world efficiently and turn it into high-quality training data.

Generative Bionics and AMD are building a Physical AI data factory Generative Bionics is partnering with AMD to build a 5,000㎡ Physical AI Data Factory in Italy. The first facility is scheduled to begin operations in Q4 2027. The goal is to connect real-world robot operations, data collection, digital twin simulation, and model training in one pipeline. GENE.01 humanoid robots will collect multimodal data while performing industrial tasks. Digital twins will expand the range of training scenarios, while AMD provides the computing infrastructure for data processing, World Action Model training, and large-scale simulation. Trained models will then be tested on physical robots. --- behind this move... NVIDIA has been building a broad Physical AI ecosystem spanning robotics development, simulation, foundation models, and edge computing through Isaac, Cosmos, GR00T, and Jetson. AMD needs more real-world projects to demonstrate that its computing platform can support the entire robotics development pipeline. This partnership also gives AMD a foothold in two markets: data center computing for model training and simulation, and onboard computing for AI inference and real-time robot control. Every new round of data processing, model iteration, and robot deployment could generate additional demand for computing power.

Imagine AMRs moving parts through a factory while humanoid robots handle components or tend machines. In a shared workspace, a robot needs to make sense of more than what’s in front of it. It has to connect what it sees now with what it saw a moment ago, and account for what has moved. For #RopediaSuperWeek, @ropedia_ai’s S-Agent 2.0 explores this through visual understanding, precise measurements, and persistent scene memory. It scored 66.5 on MMSI-Bench and 82.3 on ReVSI, outperforming existing agentic systems across all five spatial reasoning benchmarks evaluated.

Samsung is testing a humanoid designed for 24/7 operations. The bipedal robot is being trained for navigation and material handling at Samsung’s R&D center in Seoul, with staff and partners monitoring its operation in shifts. (see a bit in the referenced video.) Samsung is considering a CES 2027 showcase... Keeping robots working around the clock means minimizing downtime. Boston Dynamics’ Atlas and UBTECH’s Walker S2 use autonomous battery swapping, with the process taking around three minutes. Agility’s Digit uses fast charging, Figure 03 supports wireless charging, and Tesla has demonstrated Optimus autonomously docking to recharge. Each approach aims to maximize productive hours. Samsung is also advancing its robotics AI. Its Metro-WM world model uses previously observed states to guide robot planning. In specific long-horizon task tests, it improved success rates by up to 37.33 percentage points and made planning up to 10.9 times faster. and it is also developing dexterous hands for tasks such as grasping, assembling and connecting flexible wires. These remain research advances, with no confirmation that they are already integrated into the humanoid under testing. Samsung has a major advantage in its existing industrial footprint across semiconductors, displays and home appliances. It can test humanoids in its own factories, collect real-world operational data, use that experience to improve its models, and expand into more tasks and deployment scenarios.