Figure shipped its third-generation humanoid in February. Tesla followed with Optimus Gen 3 in April. Both are now in commercial pilots — not factory floors, but supervised, narrow-task deployments at logistics centres, automotive lines, and one quietly impressive hospital.
The demos are unfair. The marketing footage shows a humanoid pouring coffee, folding laundry, and walking up stairs. The pilots we visited show robots doing one task — usually pallet transfer or sub-assembly handoff — in a tightly constrained cell, with a human supervisor and a clear stop button.
Where it actually works in 2026: tasks that require human-form-factor mobility (climbing onto a vehicle chassis, threading through narrow aisles) combined with low-variance manipulation. Tasks where the cost of a fixed-purpose robot exceeds the cost of a humanoid by enough to justify the lower throughput. There aren't many of those, but there are some, and the count is rising.
Where it doesn't work: anything with high-variance manipulation (consumer goods picking with packaging variance is still rough), anything safety-critical without a cage, anything requiring 24/7 uptime. The batteries last four to six hours and the dexterity-to-cost ratio is still worse than a fixed gripper for most pick tasks.
The strategic question for clients isn't 'should we buy a humanoid.' It's 'are any of our floor processes constrained by the human-shaped-mobility requirement, and would solving that unlock a downstream automation?' For most warehouses, the answer is still no. For a small but growing fraction, the math is starting to work.
