Why does one joint stop before it reaches the commanded angle?

A soft stop and a hard stop feel different

Before you touch anything, establish which of the two you have. A software clamp is silent, perfectly repeatable and lands on the same number every time. A mechanical obstruction is noisy, slightly different on each attempt, and the servo audibly works against it. Confusing them sends you to opposite ends of the workshop.

What the wrap point is

A bus servo reports its position as a number on a fixed circular scale. The scale runs from one end to the other and then rolls back to the beginning, because the shaft is a circle and the count has to close. That rollover is the wrap point, and it exists at one fixed physical orientation of the output shaft, decided by the servo, not by your build.

Calibration stores where your neutral pose sits on that scale. Every angle you command afterward is that stored zero plus an offset. If the zero landed comfortably in the middle, offsets in both directions stay inside the scale and everything behaves. If the zero landed near one end, part of the range you want falls past the rollover, where the reported number stops increasing and starts again from the other extreme. Software that let a joint cross that boundary would produce a full-speed unwind in the wrong direction, so it clamps instead. The clamp is the stop you are feeling, and it is the safe behavior rather than a bug.

This is also why the problem is arm-specific rather than model-specific. The horn can be seated on the output shaft at many different rotations, and where it goes decides where mechanical neutral falls on the electrical scale. Two arms built from the same kit by two people can differ by enough that one owner never encounters this and the other loses an afternoon to it.

Fix it

1. 1
   Rule out mechanics first

   With the servo supply off, move the joint by hand through the whole range you want. If it goes there freely, nothing physical is stopping it and you are chasing a stored number, not a screw.

2. 2
   Write down where it stops, in both directions

   Note the commanded value at the stop going each way. A block in one direction only is the signature of a zero point near one end of the scale. A symmetric block is a narrow recorded sweep instead, which the same recalibration also fixes.

3. 3
   Remount the horn so neutral sits mid-scale

   Bring the joint to the pose you consider home mechanically, loosen the horn, and reseat it so the servo sits near the middle of its travel in that pose. This is the step that actually solves the problem. Everything after it is bookkeeping.

4. 4
   Recalibrate the arm

   Guide every joint to the neutral pose, confirm, and let the client capture the reference. Calibration is a whole-arm operation, so the other five joints get a new zero at the same time whether they needed one or not.

   bash

   ay-robots calibrate

5. 5
   Sweep the joint to both limits and watch

   Walk the joint slowly to each end of the range you need. Stop immediately if anything sounds strained. What you are confirming is that the range you actually use now sits inside what was recorded.

6. 6
   Check that no joint landed near an edge

   Do this after every recalibration, not only after this one. A home value sitting close to either end of the scale is the same bomb waiting for a different joint, and it costs one look to notice.

7. 7
   Home the arm and drive it

   Return to the home pose and drive the repaired joint from the client through its full intended travel. If it goes where you send it, you are finished.

   bash

   ay-robots home

# Confirm the arm is complete before recalibrating
ay-robots devices

# SERIAL PORTS
#   /dev/ttyACM0   SO-100  (6 servos responding)

# Capture the zero point: guide each joint to neutral, then press Enter
ay-robots calibrate

# Return to home and drive the joint from the client
ay-robots home

Why this often appears right after a repair

Replacing a servo, removing a horn to get at a screw, or rebuilding a joint after a collision all change where neutral sits on the electrical scale. The arm worked yesterday, you fixed something unrelated, and now one joint stops short. The two events are connected through the horn, not through whatever you were originally repairing, which is why the timing tends to send people looking in the wrong place. Full first-time assembly and calibration is on /learn/so-100-getting-started if you want the reference procedure.

What it is not

• A limit switch. These arms have none, so there is no switch to bypass or replace.
• A firmware restriction you can unlock. The clamp exists because crossing the wrap point would unwind the joint at speed.
• A policy problem. A trained policy commands positions like anything else and is clamped the same way.
• Something a longer training run improves. If the joint cannot reach, more steps teach the policy to stop there more confidently.
• A sign the servo is failing. A servo that answers, holds torque and tracks inside its allowed range is healthy.

The cost of leaving it alone

A clipped joint quietly shapes your data. During teleoperation your hand keeps moving while the joint stops, so the recorded action column flattens against the limit for the rest of that segment. The policy learns that the correct action at that point is to stop, and it reproduces that faithfully at inference, which then looks like a model that will not reach. You can spend a long time adjusting hyperparameters for a problem that lives in a horn. If you have already trained on such a dataset, the fix is the horn, then a fresh recording, then a new run: the model side of that loop is described on /train/act-on-so-100.

Frequently asked questions