Industry coverage highlights partnerships between chip and EDA leaders and robotics stacks — aimed at sim-to-real transfer: train in rich physics-aware simulation, deploy with fewer surprises on the factory floor or warehouse lane. The goal is not theatrical humanoids alone; it is dependable throughput where downtime costs real money.
World models and data
Collecting diverse embodied data — grasp failures, lighting shifts, uneven floors — remains expensive. Teams blend teleoperation, simulation rollouts, and careful curriculum design so policies generalize beyond the demo bench.
Physical AI fails loudly when physics disagrees with the training distribution — which is why instrumentation matters as much as architecture.
What we’re watching
Safer collaborative robots, better tactile feedback, and evaluation suites that measure reliability under load — not just peak benchmark scores on curated clips.