New to OmniHand 2025: how to get your first grasp working

I spent two days trying out AGIBOT's OmniHand 2025. "Trying" is a bit of a stretch—I was actually helping a friend get started from scratch. He does non-standard automation, took on a small job that needs a hand to grab material blocks off a production line. I've had the Pro version for a month and hit some pitfalls, so I went over to walk him through this 2025 from unboxing to its first successful grasp.

First, let me be clear about what this thing is. A dexterous hand is a robotic hand mounted on the end of a robotic arm, replacing humans to grab, pinch, and twist. It's different from the two-finger grippers in factories—grippers only open and close, while a dexterous hand has several fingers, each with several joints, that can bend and swing. OmniHand 2025 has 16 degrees of freedom—in plain terms, this hand has 16 independently moving joints. The tactile version has over 400 tactile force-control points across the whole hand, essentially a skin on the fingers that can sense pressure.

This post covers just one thing: how to go from unboxing to getting this hand to steadily grab a material block.

Day one: mount it first, don't rush to power on. Unbox and count everything—the hand itself, the flange adapter plate, a set of screws, the CAN-FD cable and termination resistor—don't miss a single item. What's a flange? A metal plate with a hole in the middle, screwed onto the end of the robotic arm on one side and onto the hand on the other, connecting the two.

When mounting, first align the adapter plate with the flange on the end of the robotic arm, confirm the four holes line up—if they don't, don't force it, switch hole positions. Screw in diagonally, don't tighten all the way at first—get all four started, then tighten them one by one. My friend tightened one screw all the way first, and the other three all went crooked. After the hand is mounted, check the wiring harness—the CAN-FD cable comes out from the base of the hand and should run along the robotic arm; don't let it dangle in mid-air or it'll get yanked as soon as anything moves.

Here's a pitfall I fell into the first time too. CAN-FD is the bus this hand uses to communicate with the outside—two signal lines, one high, one low. Reversed, it won't connect, but it won't burn. If it won't power up or the host computer can't connect, suspect these two wires first.

Day two we got stuck on the wiring harness; day three we officially powered on. Don't get the power-on sequence wrong: connect the emergency stop first, then power the hand, and finally open the host computer. The emergency stop is a big red button—hit it and all power to the hand cuts off. During debugging there must be one within reach.

The host computer is the control software running on a PC; the SDK from the manufacturer includes examples. Once connected you'll see an interface with sixteen joint angles updating in real time. The expected result is that when you don't touch it, the numbers are stable, with slight jitter. Find the connection in the software, select the corresponding serial or network port, click Connect—once connected the status light goes from blinking to solid. Don't send grasp commands yet—find Enable to enable the motors; the hand will click and tense up, which is normal. Then click Zero to return to the zero position—all fingers slowly open to a standard pose.

Here's the pitfall: if you send angle commands without zeroing first, the fingers will jam hard and make a continuous abnormal noise. It doesn't know where it currently is—just zero and start over. The moment you hear the jamming sound, disable immediately, zero, and start over. My friend jammed it three times in a row that day and felt awful about it.

Before actually grabbing something, there's one step you can't skip: calibration. Calibration means telling the hand which point in the real world your fingertip's current position corresponds to. I'm also new to calibration—I've been trying it these past few days, wouldn't say I'm proficient, but I've got the process figured out: touch the fingertip to several known positions in sequence, the software records them, and only then do the coordinates you send mean anything.

After calibration, grasping is actually simple. In the software, select a grasp pose—for example, an enveloping grasp, where all five fingers close to wrap around the object. Place the material block in the palm and close slowly; with the tactile version you can see the pressure values at each tactile point rising. Stop when pressure reaches a threshold—you have to tune the threshold yourself; too low and it can't grip, too high and it deforms the block.

A week later my friend's line was running stably, and the block never dropped. He told me something honest.

What really takes time is the calibration and threshold tuning up front—the grasping itself was tuned in half an hour.

Next step: don't touch fine manipulation yet—try grabbing things of different hardness, a sponge, a plastic block, a metal piece, one each, and map the tactile thresholds to the force levels. Once you can roughly tell what's in the hand from the pressure curve, then go back to insertion, screwdriving, and so on. I haven't nailed down a stable calibration solution yet—whether changing workpieces means recalibrating or relying on vision compensation, I'm still testing.