Page 3 of 3. Back to the latest

A humanoid that spars with no one at the controls, a $100 million warehouse robot raise, Samsung reorganizing to build its own humanoid, and a fresh bet on the layer beneath every robot.

A warehouse robot and a home robot aren't learning from the same data. How training data requirements shift by deployment environment, and what stays constant no matter where a robot works.

Nscale bets $3.5B of compute on Figure, a dual-arm robot ties rebar on a building site, Medtronic puts $700M into surgical robotics, and Nvidia pushes real inference to the edge.

Robot recordings capture behavior that unfolds over time. Annotation is what turns that into actions, states, events and outcomes a learning system can use.

Skild's S1 learns from one video, Figure scales human data with Index, humanoids race in Beijing, and XPENG raises to build IRON.

Simulation can accelerate robot training, but the physical world doesn't follow the rules of a simulator. Here's where simulated data works, where the reality gap appears, and why real-world data still matters.

Diversity, sensor coverage, and consistent annotation matter more than raw volume. Here's how to tell if a dataset can actually teach a robot.

From observation to operation: how robots acquire real-world skills.

Robots don't have to learn every job from scratch. They start with us.