SO-101
Course overview/Chapter 05 · Arrival Checks
中文
Chapter 05Kinematics~20 min

Arrival Checks

Companion page: the Arrival Check tab in the console.

10 sections 4 figures 2 code blocks ReplayToleranceMeasurements
Trajectories reproduced under the three criteria

1. Replay Only Does Two Things#

A recording is a series of frames: one every 1/30 s, each holding the six joint angles, the TCP pose and the gripper opening.

Replay is a loop that does two things:

text
take frame i  →  send the six joint targets  →  decide when to send frame i+1  →  …

The TCP pose is recorded too, but it is used by the "end-effector pose arrival" criterion; the actual command is joint targets.

The second of these is the arrival check. It does not change positioning accuracy; it only sets the pacing — which directly determines how long a replay takes and how faithfully the trajectory is reproduced.

2. The Three Criteria#

Criterion What decides "the next frame may be sent" Parameters Characteristics
Fixed timing No check; send on schedule Send interval (ms, default 33) Fastest; when the arm cannot keep up it "smooths away" the corners
End-effector pose arrival Measured TCP position + pitch/roll + gripper Position mm, orientation °, gripper Most accurate; also the slowest
Joint angle arrival Measured angles of the six joints Joint angle °, gripper Moderate; but joint error ≠ end-effector error

All three run through the same loop; they differ only in the wait condition.

3. Running a Comparison on This Page#

Using this page essentially means replaying the same recording several times with only the criterion changed, then comparing three things: elapsed time, how well the two curves overlap, and path length.

Order of operations

The figure above shows every control on this page; the numbers are the order of operations. The steps below follow that order.

Step 1: Set the start point#

  1. Support the arm first → click Teach → once confirmed, torque is released → move the arm by hand to the desired start point
  2. Click Record anchor → the six joint angles are saved
  3. Click Finish teaching again to restore torque

Simulation has no "drag by hand": pose the arm from another page first, then click Record anchor.

Step 2: Perform the motion to be replayed#

  1. Motion recording card → click Record motion - On the real robot, torque is released again and teaching mode starts automatically
  2. Guide the arm by hand through the entire motion (slowly — do not fling it)
  3. Click Stop recording - A dashed orange line appears in the 3D view: the path the TCP just traveled

Step 3: Return to the start point#

Click Return to anchor and wait until the message reads "Already at the anchor (deviation X°)".

Replay must start from the anchor; otherwise the first frame "teleports".

The anchor may have been recorded outside a soft limit (hand-teaching can push a joint a few degrees past one). Return to anchor is not clipped by the soft limits: it goes all the way back to the recorded pose instead of stopping on the boundary. The E-stop, the workspace rules and the speed limit still apply.

Step 4: Choose a criterion and run once#

  1. Choose the mode in the Arrival Check card - The pose and joint modes each have their own tolerances; they do not affect each other
  2. Click Start replay
  3. During replay the 3D view draws a solid green line — the trajectory actually traveled

Step 5: Read the three metrics#

Metric Where to look Notes
Elapsed time Progress text on the replay card (112 / 359 frames) The stricter the criterion, the longer it takes
How well the two curves overlap 3D view; clearest viewed from above Orange = the path it should take, green = the path it actually took
Path length View data → summary A green line clearly shorter than the orange one = cutting corners
View data

The path length in the summary is the most direct measure of whether the arm is cutting corners: the recording measured 292 mm, while the 8° joint run traveled only 228 mm.

Step 6: Re-run with a different criterion#

First click "Clear actual trajectory", or the previous green line will still be there and the old and new runs will be impossible to tell apart. Then repeat Step 4.

Want to switch criteria mid-run? Click Stop, change the setting, then Resume from stop or Restart from the beginning. Resume from stop does not require returning to the anchor, which makes it suitable for comparing the remaining segment of two criteria from the same position.

4. What the Tolerance Actually Measures#

End-effector pose arrival

Parameter Measures Default
Position tolerance (mm) Spatial distance between the measured TCP and the TCP of the recorded frame 5 mm
Orientation tolerance (°) Difference in pitch and in roll, each separately 3°
Gripper tolerance Difference in gripper opening (0–100) 3

Joint angle arrival

Parameter Measures Default
Joint angle tolerance (°) Difference of each of the six joints from the recorded frame 3°
Gripper tolerance As above 3

Tolerance is the threshold at which the system considers the target reached; it is not an accuracy figure. It is a different thing from the arm's real positioning accuracy — setting the threshold below that real accuracy only makes every frame wait until it times out.

5. Measurements: One Recording, Six Configurations#

A 359-frame recording (12.0 s, TCP travel 292 mm) replayed once under each of six configurations:

Configuration Replay time Max deviation Mean deviation Actual path length
Fixed time 33 ms 14.0 s 4.5 mm 0.11 mm 300 mm
Fixed time 66 ms 24.0 s 4.2 mm 0.09 mm 300 mm
Joint angle 1° 17.7 s 3.1 mm 0.23 mm 287 mm
Joint angle 8° 7.1 s 12.0 mm 4.72 mm 228 mm
End-effector pose 2 mm 21.2 s 0.85 mm 0.08 mm 287 mm
End-effector pose 15 mm 10.2 s 4.7 mm 1.60 mm 280 mm

("Deviation" = shortest distance from each point on the actual trajectory to the recorded trajectory; "path length" = total distance actually traveled.)

Time and deviation

Four Things the Table Shows#

① The fastest is the crudest. A joint tolerance of 8° took only 7.1 s (40% faster than the recording), at the cost of a 12 mm maximum deviation and a path length of only 228 mm — 22% shorter than the recorded 292 mm. This is not "moving fast"; it is cutting corners: every turn gets clipped.

② The most accurate is the slowest. Pose 2 mm took 21.2 s (1.8× the recording), but its maximum deviation was only 0.85 mm.

③ In timed mode the error concentrates at the corners. The 33 ms setting has a mean deviation of only 0.11 mm, which looks very accurate, but a maximum of 4.5 mm — the arm keeps up on straight segments and falls behind in sharp turns, overshooting a little and then coming back.

④ Joint tolerance and end-effector error are not the same thing. At a joint tolerance of 1° the maximum deviation was 3.1 mm; yet the same 1° can correspond to 5–6 mm with the arm fully extended. Converting a fixed angular error in joint space into an end-effector distance depends on the current pose (the Jacobian) — this is the most fundamental difference between "joint arrival" and "pose arrival".

INTERACTIVE LAB

Replay comparator

Measured data from Chapter 5

One recording — 359 frames, 12.0 s, TCP travel 292 mm — replayed once under each of six configurations. Pick a configuration to compare replay time, maximum deviation and actual path length.

Current configurationFixed time 33 ms
Replay time14.0 sRecording baseline 12.0 s
Max deviation4.5 mmSmaller is closer
Actual path length300 mmRecording baseline 292 mm
The data comes from the measurement table in §5 of Chapter 5; this lab only selects and compares, it does not refit anything. Read path length together with deviation: a much shorter path usually means cutting corners.

6. Why Joint Arrival Cuts Corners#

Because it only requires "each of the six joints is close to its target" and does not care where the arm actually is.

An extreme example: with the arm fully extended, 1° at the shoulder joint corresponds to roughly 5 mm at the end effector. A tolerance of 8° allows every frame to be off by 40 mm — so the arm simply skips the intermediate frames and heads straight for the end of the next segment. Rounded corners become sharp corners, and curves become polylines (the middle panel of the figure at the top of this chapter).

This is also why trajectory fidelity (gluing, insertion, surface following) requires the pose criterion; the joint criterion suits only cases where "roughly the right motion" is enough.

7. How Arm Accuracy and Control Rate Affect the Choice#

Comparing three classes of arm#

Smallest controllable angle Repeatability Typical tolerance setting
Industrial 6-axis (UR / Fanuc class) ~0.001° (high gear ratio + absolute encoders) ±0.02–0.1 mm 0.1 mm / 0.1°
Piper (lightweight cobot) 0.13° Sub-millimeter 0.5 mm / 0.5°
SO-101 (this unit) 0.088° (4096 counts/rev) millimeter level 1–5 mm / 1–3°

Note: a fine minimum step does not mean high accuracy#

The SO-101 encoder is 4096 counts/rev, 360 ÷ 4096 ≈ 0.088° — finer than Piper's 0.13°. But that only says how finely the encoder can count, not how accurately the arm can stop. What actually determines accuracy:

  • Backlash in the plastic gears: the joint has to turn a few degrees in the reverse direction before anything moves
  • Stiffness of the 3D-printed structure: the end effector deforms as soon as a force is applied
  • No gravity compensation: the arm sags a little when torque is applied or released
  • Calibration error: the mid-range position is confirmed by hand

So this arm's repeatability is at the millimeter level, not the hundredth-of-a-millimeter level that 0.088° suggests. 0.088° is resolution; millimeter level is accuracy — two orders of magnitude apart.

The control-rate constraint#

The control loop runs at 30 Hz, so it only gets to look at the current state once every 33 ms:

  • Smallest resolvable motion: at 60 °/s the joint has already moved 2° within 33 ms. With a tolerance below 1°, the verdict flips back and forth between consecutive frames.
  • Check latency: even with a sensible tolerance, confirming arrival costs one extra cycle.

What happens when the tolerance is too tight#

Each frame's wait has a 3 s timeout. With a tolerance tighter than the mechanical accuracy, every frame waits the full 3 s before giving up:

text
359 frames × 3 s ≈ 18 min

And after "giving up" the arm is not actually in position anyway — it is no more accurate, just tens of times slower.

Conclusion: for a low-accuracy arm, prefer fixed-timing commands#

Arm Recommendation
Low-accuracy arms such as the SO-101 Fixed timing first
Lightweight cobots such as Piper Fixed timing as the default; pose 1 mm when needed
Industrial arms End-effector pose arrival is meaningful; worth waiting for 0.1 mm

Why fixed timing is the default for low-accuracy arms:

  1. The accuracy bought by an arrival criterion is eaten by mechanical error. Pose 2 mm does push the maximum deviation down to 0.85 mm, but this arm's repeatability is itself at the millimeter level — chasing 1 mm has no practical meaning, and it costs 51% more time (14.0 s → 21.2 s).
  2. The mean deviation is already very small. Fixed timing at 33 ms gives a mean deviation of 0.11 mm, among the best of the six configurations (66 ms is 0.09 mm but slower overall). The error is concentrated in the instants of sharp turns, where a low-accuracy arm is poorly controlled anyway.
  3. It never stalls on "never arriving". Fixed timing does not depend on measured feedback, so no tolerance setting can make it time out.
  4. It works predictably with the speed slider. Slow speed plus fixed timing makes both trajectory smoothness and duration easier to predict.

When higher fidelity is needed, fall back to joint arrival at 2–3° rather than fighting for a millimeter-level tolerance.

8. Choosing a Criterion#

Scenario Recommendation
Running a full motion on a low-accuracy arm (SO-101) Fixed timing 33–66 ms
Just checking whether the motion is right Fixed timing 33–66 ms — the fastest option
Trajectory-insensitive motions such as pick and place Joint angle 3–5°
Gluing, insertion and surface-following motions (high-accuracy arms) End-effector pose 2–5 mm, orientation 2–3°
First run on the real robot Start with fixed timing 66 ms to watch the whole motion, then switch criteria
Replay keeps stalling The tolerance is too tight; loosen it first

Also do not forget the speed slider: it limits how fast the target advances, and it is an independent knob from the arrival check — one governs "how fast it moves", the other "when it counts as arrived".

9. Common Misconceptions#

Claim Correct? Explanation
"A smaller tolerance is more accurate" ✗ Below the mechanical accuracy it is meaningless, and timeouts make replay unusably slow
"High encoder resolution means high accuracy" ✗ The SO-101 resolves 0.088°, yet its repeatability is only at the millimeter level
"Joint arrival equals end-effector arrival" ✗ The same 1° maps to end-effector distances that differ several-fold between poses
"Timed mode has no accuracy" ≈ The mean deviation is small; the error concentrates at the corners
"The arrival check improves positioning accuracy" ✗ It only decides when the next frame is sent; it does not change the servos' own accuracy
"Parameters tuned in simulation transfer straight to the real robot" ✗ The real robot has backlash, sag and noise; loosen the tolerances by 2–5×

References#

  • Chapter 3: End-Effector Pose — the five components of a pose, joint and line paths
  • Chapter 4: FK and IK — why converting joint error into end-effector error depends on the pose
  • Experimental data: docs/data/arrive-experiment.json; figure script: scripts/make_arrive_figures.py