
How Can a Beginner Tune This Robot Squat-and-Stand Motion from Scratch?
Last week I had dinner with a team doing embodied AI. I asked them how long it takes to choreograph a bending-down-to-pick-up motion. The engineer thought for three seconds and said, two people, three days.
I didn't quite believe it, so I went back and opened an ODance Studio account myself, and from scratch choreographed a squat-then-stand motion, spending about two evenings.
Not hard, but the pitfalls are real. This piece is a walkthrough for someone who's never touched it.
First, what this tool is for. ODance Studio by Oak Qingxi Technology, classified under humanoid robot software, does four things: pose the robot, let the software fill in the in-between, adjust waist and feet separately, and run it to see if anything goes wrong. "Motion control" sounds intimidating, but it just means managing how the robot moves, where it moves to, and how fast.
From importing a model to running your first motion
Step one is getting the model file for the robot you want to choreograph. This is usually a description file stating how many joints the robot has, how many degrees each joint can rotate, and how long each bone is. Think of it as a body manual. There's a common format in the industry, usually provided by the hardware manufacturer. If you don't have one, go ask the company that gave you the robot—you can't skip this step, don't think about drawing it yourself.
After importing, the interface is roughly three sections. The left column is the joint tree, one row per joint, expandable to see hierarchy; the middle is the 3D view, where your robot stands; at the bottom is a horizontal timeline marked in seconds. Layout may differ slightly by version, but the logic holds.
1. Drag the timeline cursor to 0 seconds, pose the robot in the 3D view into the starting posture, click Add Keyframe.
2. Drag the cursor to about 3 seconds, pose the second posture, say knees bent halfway, click Add Keyframe again.
3. Drag to 6 seconds, pose standing straight, click a third time.
4. Click One-Click Trajectory Interpolation.
"Keyframe" is confusing the first time you hear it. It's what the robot should look like at that second. You only mark a few key poses, and the software calculates the transitions. Interpolation is the software filling in every frame between two poses. This used to be done manually frame by frame, and one motion took two days minimum.
My habit is to start a motion with three to five keyframes; more makes it harder to tune. Interpolation usually offers several options—beginners should use the default smooth, don't jump straight to linear.
After filling in the in-between frames, go through waist and feet separately. This is the biggest difference between humanoids and robotic arms. Waist pitch and rotation determine where the center of gravity falls; foot placement determines how large the support area is. The platform has dedicated panels for these two. Beginners most easily produce motions that look like they can stand but actually can't at this step.
Finally click Simulate to run it. The built-in risk detection flags problem areas, mainly three types: joints exceeding limits, self-collision, and center of gravity leaving the support area. Run it before moving—don't skip this step.
Three places I got stuck
First was joint limits. In the 3D view you can pose the arm at any angle, but that doesn't mean this robot can do it. The first time I forced a wrist angle, simulation errored out immediately. I learned to run risk detection after each keyframe instead of saving it all for the end.
Second was interpolation time too short. Two keyframes too close together make the software fill in absurdly high angular velocity, and the robot jerks in simulation. The fix is to spread the two keyframes apart, or add a transition frame in between. From my testing, giving a simple motion five seconds or more is more stable.
Third was mixing up coordinate systems. The same joint angle, using absolute angle versus angle relative to the previous frame, produces completely different motions. I got it backwards the first time—meant to squat down, ended up leaning backward. This can only be caught by running simulation multiple times.
One more thing that's not a pitfall but must be said: passing simulation doesn't mean it runs on the real machine. Simulation has no ground friction, no load, and no motor response delay. Before going on the real robot, cut the speed to half and test first.
After learning this, don't rush into complex motions next. First take your robot's most common motion—like bending to pick something up or turning in place—break it into three to five keyframes and reproduce it, pass risk detection, then consider going on the real machine. After that, go ask the team for their own motion library, see how others distribute keyframes and set interpolation parameters—that's more useful than reading docs.
I wrote a piece a while back on hands-on impressions of humanoid hardware; progress on the hardware side is visible. But recently looking at several early embodied AI teams, everyone's talking about hardware and large models, while motion development efficiency is underestimated. One motion takes three days at a client site, and delivery costs can't come down. The tool layer's ceiling may not be high, but it's stuck in the middle of the whole chain—if it's not solved, you can't get through.
Physix Frontier