Why Agility Training Means Smarter Bodies, Not Just Cheaper Hardware

The Parcel-Sorting Showdown

In a recent live stream, a Chinese robotics company called Zibian (自变量) put a two-armed robot in front of a conveyor belt. The robot had only standard grippers—no dexterous hands, no legs. Its job: sort a stream of random parcels—boxes, soft bags, cylinders, foam-wrapped food—varying in size, weight, and orientation. Over one hour, it processed 1,816 parcels. That's 45% faster than Figure AI's humanoid, which had set a benchmark of 1,248 parcels per hour over a 200-hour run. The accuracy? Over 98%. The cost of Zibian's hardware? Roughly 70% less than Figure's.

This isn't just a robot race. It's a lesson in agility—not the kind you see in a gym, but the kind that matters when a machine has to handle the messy, unpredictable real world. The robot didn't win by being more human. It won by being smarter about how it used its body.

What Agility Really Means

Agility is often mistaken for having more moving parts. More joints, more sensors, more degrees of freedom. But agility is really about the ability to adapt—to read a situation, choose the right action, and execute it smoothly. In robots, that's a combination of perception, prediction, and control. And sometimes, less hardware can be more agile if the brain is good enough.

Take the parcel-sorting task. The robot had to identify each package's size, shape, material, and orientation. It had to decide where to grab, how much force to use, and what path to move. Some parcels had labels facing the wrong way, so it had to flip them. All this in real time, without human help, for a full hour.

Hardware Minimalism, Model Maximalism

Zibian's approach is to keep the hardware simple and push complexity into the software. Their model, WALL-B, uses a unified architecture that combines vision, language, action, and physical prediction into a single network. It doesn't just decide what to do; it predicts what will happen next. For example, when grabbing a soft bag, it anticipates whether the bag might slip. When pushing a box sideways, it predicts whether the box will slide, rotate, or topple.

This predictive power lets the robot use its two grippers creatively. For a small, light bag, it grabs and moves quickly. For a large box, it switches to two-handed support or pushes from the side. To adjust a misaligned label, it might flip a small item using inertia, or nudge a heavy box in steps. It can even spread out a soft clothing bag to flatten it before scanning the label.

The Cost of Extra Limbs

Figure AI's humanoid has a full body and five-fingered hands. That's impressive, but it comes with costs. Each joint, actuator, and sensor adds expense, complexity, and potential failure points. In a 24/7 warehouse, every extra part is a liability. Zibian's robots, by contrast, use off-the-shelf industrial grippers. They're cheaper, tougher, and easier to maintain. The missing dexterity is compensated by the model's ability to improvise.

This is a trade-off. Figure keeps the door open for tasks that require human-like manipulation—using tools, opening doors, fine assembly. Zibian focuses on the job at hand: sorting parcels. For that, legs and fingers are just overhead. As they put it, a warehouse doesn't need to pay for walking and ten fingers.

From Homes to Warehouses

Zibian actually started in the home. In 2024, they released WALL-A, their first end-to-end model. By 2026, they had robots in real homes, doing chores like folding towels, tidying desks, and cleaning. They even partnered with a domestic service company to offer robot-assisted cleaning visits. That's a tough test—homes are unpredictable, with objects scattered and environments changing daily.

The same model that learned to handle socks and cereal boxes is now sorting parcels. The underlying skills—recognizing objects, understanding spatial relationships, predicting physical outcomes—transfer across domains. The model is general; the body is task-specific. A home robot might use a dexterous hand, while a warehouse robot uses a simple gripper. This reuse cuts development costs for new applications dramatically. Instead of retraining from scratch for each new line, the model brings prior knowledge, so adaptation is mostly about hardware and workflow.

The Bottom Line

In the end, agility isn't about looking human or having the most degrees of freedom. It's about achieving the task efficiently and reliably. The metrics that matter are efficiency, cost, stability, and scalability. Zibian's parcel-sorting demo shows that a cheaper, simpler robot, powered by a smart model, can outperform a more human-like one in a specific industrial task.

This is reminiscent of what DeepSeek did for large language models—delivering strong performance at a fraction of the cost. Zibian is doing the same for physical AI. The future of robotics might not be more humanoid, but more practical—a partner that does the job well without breaking the bank.