By Sarah Xiao 
Walk the halls of CIIF 2026 in Shanghai and the most telling absence is the one nobody mentions: bipedal robots. Of the 400-plus robot exhibitors and 300-plus product launches, coverage from Gaogong Robot (高工机器人) counted almost no double-legged platforms on the floor. What filled the space instead was more mundane and more important — wheeled humanoids and dual-arm workcells actually running production tasks: shoe-sole gluing, mixed-line palletizing, screwdriving, stamping machine tending.
The strongest signal of the week, though, came off the floor: FAIRINO (法奥机器人)， the Suzhou-based cobot maker, closed consecutive C+ and D rounds totaling close to RMB 1 billion (roughly US$140 million), led by Meituan Longzhu and Meituan. Investors rarely fund a hardware category at that size twice in one quarter unless they believe the product form has settled. Their read: the near-term product of industrial embodied AI is not a walking humanoid. It is a force-controlled arm that can hold a torque window — sold as a working cell.
Key takeaways
- Biped demos lost the stage. Wheeled humanoids and force-controlled dual arms carried the real production demos at CIIF 2026.
- Force control became a spec-sheet row. FAIRINO’s ART series integrates torque sensors into all seven joints, claiming ≤0.3 N·m sensing accuracy and a 1 kHz closed loop, with 3–14 kg payloads and repeatability better than ±0.03 mm.
- The unit of sale moved from “robot” to “workcell.” Glue dispensing, box folding and welding cells now ship with vision, process parameters, end effectors and data traceability bundled in.
- A RMB 1B vote on standardization. The round funds core-component capacity — and FAIRINO says it has consolidated its harmonic reducer SKUs to ten universal models across the ART line.
- Buyers should recalibrate diligence. “Force control” is now a marketing word. The verification checklist at the end of this article is what will protect your budget.
What actually changed on the show floor
Coverage of Gaogong Robot‘s show floor focused on FAIRINO’s stand, which skipped the token walking humanoid in favor of cells doing real work: shoe-upper gluing and sole pressing, mixed-SKU palletizing, screw fastening, mixed-part welding, and stamping press tending. None of these are glamorous. All of them expose the failure modes that actually stop robots from shipping: contact force that drifts, incoming parts that don’t match CAD, tool changes that don’t fit, vision that misses deformed workpieces.
Two products anchored the story:
ART series industrial force-controlled arms. Each of the seven joints carries an integrated torque sensor; the company claims ≤0.3 N·m-class force-sensing accuracy with a 1 kHz closed-loop response, payloads from 3 kg to 14 kg, and repeatability better than ±0.03 mm. That combination targets the classic cobot gap: assembly and fastening tasks where the robot must feel contact instead of just reaching a position.
BW10-Lite, a wheeled dual-arm platform built with Boyin Hechuang (博银合创)， pairing the ART arms with Boyin’s “Bozhi-Bolt” industrial vertical model. The interesting detail reported from the stand was not the spec — it was fleet behavior: multiple units negotiate task allocation based on material, tote and partner status, and correct positions at hand-off.
Vendor figures on the shoe-gluing cell (developed with Yingshen Intelligence, 影身智能)： a 4D-native world model running on commodity RGB cameras instead of depth sensors, ±0.02 mm positioning on glue trajectories, 1.5× per-station efficiency, 99.9% yield, half-day commissioning and a 1.5–2 year payback. Treat these as vendor claims pending independent benchmarks — but the architecture choice (RGB over depth cameras) is the part with real cost implications.
The force-control spec race — and what to verify
Once one vendor puts torque sensing in every joint, the claim spreads fast. Here is the gap between marketing language and the numbers that matter:
| Marketing claim | What it should mean | What to verify |
|---|---|---|
| “Force-controlled” | Joint-level torque sensing | Is the sensor at the joint, the wrist, or the controller estimating current? |
| “≤0.3 N·m accuracy” | Detectable force change at the joint | Measured at which payload, which temperature, after how many hours? |
| “1 kHz loop” | Control bandwidth | Response latency, not just sampling rate — ask for step-response data |
| “±0.03 mm repeatability” | Pose repeatability | With full rated payload and at workspace corners, per ISO 9283 |
| “AI process model included” | Skill generalization | Trained on your product family or a demo piece? Retraining cost? |
From arms to workcells: the unit of sale moved
The quieter story at CIIF was commercial. Vendors have stopped selling general-purpose arms for these stations. The shoe-gluing cell, the dual-arm box-folding cell and the AIRLab welding agent (aimed at shipyard panel welding, with weld-seam recognition, point-cloud reconstruction and path planning) all ship as packaged units: process parameters, force-control strategy, vision, end-effector mounting and data traceability in one quote.
For buyers this changes what you are actually evaluating. You are no longer comparing robot payload curves; you are comparing how quickly a packaged cell reaches your product family, and what happens when your product family changes. The relevant questions shift from “what is the repeatability?” to “what does retraining cost, who owns the model, and can I export my process data?”
The base-layer play behind the RMB 1B round
FAIRINO’s funding is explicitly aimed at core-component capacity: the company says it self-manufactures controllers, encoders, servo drives, harmonic reducers, servo motors and brakes. Two numbers in its disclosure matter more to the supply chain than the headline:
- Ten harmonic reducer SKUs across the entire ART line — aggressive standardization that compresses cost and lead time, and simplifies spare-part stocking for buyers.
- 300,000+ units/year of stated capacity — scale that only pays off if force-controlled arms actually reach the volumes its strategic agreements suggest: 2,000 units each with Boyin Hechuang and Jack Sewing Machine (杰克科技)， the garment-machinery leader.
The strategic context: per GGII data cited in company materials, FAIRINO ranked first in China cobot brand sales and overseas shipments in 2025, expects about 18,000 global orders in 2026 (+68%), and derives 46% of overseas sales from Europe, 16% from Japan and 13% from North America, with 50+ overseas sales points.
One caution on the numbers: the 4,000-unit arm agreements are strategic cooperation deals, not confirmed shipments, and the ~RMB 1B figure is cumulative across two rounds. Both are directionally meaningful, neither is audited revenue.
A buyer’s checklist before signing a force-control order
- Ask for step-response torque data, not a “force control” badge. You want settling time and overshoot at your payload.
- Verify repeatability per ISO 9283 with full rated payload at workspace extremes — spec sheets quote no-load numbers.
- Pin down what “AI included” means contractually: who owns the trained model, what retraining costs when your product changes, and whether your process data leaves your plant.
- Price the spare-part strategy: with ten universal reducer SKUs, ask for a stocked-spares list and lead-time commitments in the SLA.
- Visit a reference site in your industry, not the vendor’s demo center — the half-day commissioning claim should be tested on a line with real incoming-part variance.
- Check the end-effector ecosystem: quick-changers, grippers and tool interfaces that match your existing tooling inventory.
What this means for component sourcing
If force-controlled arms become the 2026–2027 volume product, the demand mix shifts accordingly. Joint-level torque sensors move from “optional precision part” to every-axis BOM line — watch capacity and pricing among sensor suppliers. High-bandwidth servo drives and encoders ride the same wave. Harmonic reducer demand consolidates around fewer, standardized models, which favors suppliers who can commit to those SKUs at scale and squeezes makers of bespoke variants.
The counterpart risk is standardization itself: ten universal reducer models mean fewer differentiated purchase points and more substitutable relationships. Component suppliers to this segment should expect customers to push spec-sheet parity harder and to negotiate on data, delivery and service rather than on exotic parameters.
The industry, in short, has stopped voting on which humanoid walks best. It is voting — with factory budgets and now with a billion RMB — on which arm holds its torque window. That is a much easier product to buy, and a much harder one to fake.
Related reading
- Reading a humanoid robot spec sheet: DOF, payload and torque — how to check the force-control numbers a vendor quotes.
- Humanoid robot market H1 2026: 36,000 units, top five all Chinese — where the volume behind these launches is coming from.
- Robot tactile sensors compared: flexible arrays, magnetic Hall, vision-based — the sensing stack a force-controlled arm depends on.