
Zhongke Xinsong's Wheeled Humanoid: Force Control Delivers, But Retooling Remains the Old Problem
Last Wednesday at a client site, a small assembly line, they wanted the robot to insert a connector. The engineer showing me around didn't complain that the robot was inaccurate—he complained that every time they change models they have to re-teach the data, and teaching takes two days.
I've heard that line too many times. So these past two months I went through the whole SIASUN setup—site visits, requested videos, pored over spec sheets. First, let me clarify what it is.
BR-MR73A is a wheeled humanoid robot: wheeled base, humanoid upper body, two arms. "Humanoid" means human-like in form, not that it walks. Seven axes means seven joint degrees of freedom—one more redundant DOF than the common six-axis, which helps it get around dead angles where joints can't compute in certain postures. The BR-R73 arm focuses on bionic force control, sensing contact force, for insertion and grinding work that needs feel.
Force control really does save effort. I watched at the client site for about forty minutes, and sales sent two more operation videos. Grasping, placing, screw-driving—those segments were smooth. The biggest advantage of the wheeled base is that walking is bypassed entirely—no need to worry about balance, stability is an order of magnitude higher than bipedal. I say this with some confidence; I've seen too many bipedal falling scenes, and I wrote a piece a while back called "The Robot Has a Brain, But Still No Hands" about exactly this.
Where it gets stuck is one thing: changing workpieces. Teach a grasp point once, switch to material with a 2mm size difference, re-teach. The engineer's exact words: teaching takes two days.
Real-machine data collection is expensive, human video data has low information content, and simulated synthetic data has a Sim2Real Gap—none of these three problems has been solved.
This is from that embodied intelligence community technical guide, and I agree. Sim2Real Gap means no matter how well you train in simulation, it falls apart when moved to the real machine.
| What I saw in the demo | What I actually care about |
|---|---|
| Grasping, placing, screw-driving | How long does changeover take |
| Force-controlled insertion without hard ramming | Repeatability, long-term drift |
| Wheeled base moves stably | How to interface with existing production line systems |
| Single unit runs | Ten running together—who maintains them |
Conclusion: it depends. Suitable for two types of teams: first, production lines with fixed stations and product models unchanged for half a year—like certain 3C fixed stations—where this arm plus integrated joint modules is cheaper than imported solutions, and that's solid; second, research and data collection teams, where seven-axis force control can serve as a collection platform for contact tasks. Not suitable for whom? Flexible lines that change product models weekly—don't touch it, changeover costs will eat you alive; those expecting it to be general-purpose—also don't touch it, no robot today is general-purpose.
Next I'll watch two things: first, whether the changeover calibration process can be compressed; second, cluster maintenance costs above ten units. I predict that in the next twelve months, the marketing language for this batch of wheeled humanoids will shift from general-purpose to dedicated stations. Whoever first compresses changeover time from two days to two hours is the one who gets orders. This wave of embodied intelligence ultimately comes down to changeover cost.
Physix Frontier