BofA Analysts' Field Research: China's Humanoid Robots Pivot to Commercial Deployment
Nashnova编辑部
BofA analyst Ming Hsun Lee attended the 2026 World Robot Conference in Beijing and told clients China's humanoid-robot industry has moved from R&D showcases to commercial deployment — logistics is the first real use case to clear the bar, signaling the narrative has shifted from 'can it work' to 'can it make money.'
Why is logistics the first scenario to break through?
X Square Robot's dual-arm sorting system processes over 1,800 parcels per hour; human workers handle roughly 1,200–1,400. This means → in this scenario, the robot is not replacing human labor — it is outperforming it.
Galaxea AI's retail sorting and delivery system hit a success rate above 95%. Galbot, Spirit AI, and others demonstrated long-duration tasks — folding clothes, meal prep, picking supermarket orders.
In plain terms = logistics lands first because the motions are repetitive, the environment is controlled, and efficiency is directly measurable — it is the easiest place to prove the math works.
Why isn't the general-purpose brain ready yet?
More companies are adopting the World Action Model (WAM) — an AI architecture that lets robots understand physical environments and plan actions autonomously — aiming for a single foundation model that controls wheeled, bipedal, and other form factors.
The bottleneck is data: Lee estimates a truly generalizable model needs 10 million to 100 million hours of real-world data. Leading firms currently target over 1 million hours by end of 2026 — a gap of at least 10×.
The industry's workaround splits into two stages: pre-train on cheap, large-scale egocentric data — video captured from the robot's own viewpoint — then fine-tune on high-quality teleoperation and failure-recovery trajectories. This reflects a "feed it rough grain first, refine later" approach rather than waiting for perfect data.
How much further can component costs fall?
6D torque sensors, tactile sensors, and planetary roller screws — precision parts that convert rotational motion into linear motion — saw the fastest cost declines over the past year, driven by scale and manufacturing improvements.
Tactile sensors are now standard in dexterous hands — the robot's "fingers" — accounting for 20%–25% of the hand's bill-of-materials cost. Leading vendors' hands approach 1 million cycle lifespans.
Lee forecasts these trends can support a cumulative ~45% decline in full-unit humanoid-robot costs between 2026 and 2030. This means → if the forecast holds, a humanoid robot in 2030 will cost roughly half of today's price.
Has the dexterous-hand tech converged on a winner?
Hands shown at the conference ranged from 6 to 37 degrees of freedom: high-DOF models (approaching human-hand dexterity) target R&D and precision tasks; 6-DOF hands and grippers are widely used in manufacturing.
Tendon-driven, linkage, and direct-drive architectures are all still competing — no convergence yet. In plain terms = the industry has not picked a "standard answer"; every vendor is betting its own design will win.
This signals that dexterous hands remain in a multi-track horse race, with no single dominant standard likely in the near term.
What will prove or break this commercialization narrative?
Lee flags two checkpoints: first, whether the cost curve delivers the projected 45% decline; second, whether retail and home scenarios beyond logistics produce scaled orders before end of 2026.
This means → year-end 2026 is the first real validation window — if only logistics holds up, the "broad commercialization" narrative loses much of its credibility.
In plain terms = a demo on a conference floor is not a sale in the market. Order data over the next six months matters more than any booth showcase.
Content is for reference only, not financial advice.