Explore

Investigation

Once assembled, you realize that the robot is not the most accurate. How would you investigate this? How would you fix it?

Break manipulation inaccuracy sources into:

Control: Is the arm following the reference it is given in joint space? How does that error translate to task space? These are errors that can be measured by the system’s software, and usually point to control issues or actuation issues.

Kinematics and Taring: Do the forward kinematics match the real arm? These errors represent a disconnection between what the system’s software believes the hardware is doing and what it is doing in reality.

Perception: Is the target correctly aligned to the real world? These errors represent an inability to provide the robot arm with the correct target

Inaccuracy Sources

Inaccuracy Sources

Name

Category

Measurement

Notes

Possible causes

Motor drivers following their references

Measure joint-level error signal in driver telemetry data

"Following their reference” means that reproted values are within tolerance of reference values, not considering. Reported values differing from real-world values is out of scope

Trajectory controller following it’s reference

Measure joint-level positional error signal in controller telemetry data

Joint positions correctly tared

Use a hardware rig that fixes the passive arm in the zero-position, check telemetry reads zero within tolerance

Joint torques correctly tared

Use a hardware rig that braces each joint, check for zero-torque within tolerance.

For each joint apply torque that will apply a known force against the rig’s contact position for that joint/link, check calculated torque against measured.

Kinematic model aligns with real robot

Use a rig that attaches a 3D camera to the base with a known transform, measure a number of arm configurations, compare point cloud to forward kinematics.

Alternatively do so with a rig that provides known EEF targets and measure accuracy at touching the targets

Camera calibration

Use the checkerboard calibration technique and see if significantly differing values would be applied

Hand eye calibration

Place an april tag on the end effector mounted with a known transofrm from EEF frame, detect the pose from the camera, compare to the pose from camera to tag through the camera→base and forward kinematics

Target segmentation aligns with object in camera image

Check recorded data of inaccurate pick tasks, visually validate the segmentation placement on the images used for pose estimation.

Target pose aligns with object in camera frame

Use a 3D visualizer to check pose placement relative to image and known conveyor height.

There are no rows in this table

⁠

Lower Level

Lower Level

Name

Test

How to fix

Bad encoder

Measure encoder readings against known joint displacements

Replace

Bad torque measurement

Apply known torque and measure output

Replace

Infeasible references

Assess velocity and acceleration values of the references passed to the driver and compare to motor specs

Adjust constraints in motion planning/control

Bad dynamic model

Perform isolated joint-link system identification and compare outputs to full model parameters

Perform more intensive system identification

Unmodeled hardware dynamics

Collect data over a wide range of motion, compare model estimates to measured values, drill down to specific areas of disagreement.

Look for patterns and signals of things such as joint deformities, high joint friction, irregular joint axes

Increase specificity of dynamic model to account for unmodeled dynamics

Encoder drift or hysteresis

Endurance test of repeatable displacements

Replace sensors with less drift/hysteresis or put in place a monitoring system that can alert to the need of taring or parameter adjustment

Torque sensor drift or hysteresis

Endurance test of repeatable torques

Replace sensors with less drift/hysteresis or put in place a monitoring system that can alert to the need of taring or parameter adjustment

Incorrect joint axis placements

Perform measurements per joint at right angle configuration to validate axis placements

Use outputs of left-mentioned procedure to update joint axis placements

Malfunctioning actuator

Feasible references do not produce expected output

Replace

Improperly conducted camera calibration

Re-running procedure multiple times produces consistent results different to the applied results

Use outputs of re-run

Improperly conducted hand-eye calibration

Re-running procedure multiple times produces consistent results different to the applied results

Use outputs of re-run

Local minima in hand-eye calibration

Run multiple times on same calibration data tuning parameters and checking for jumps in converged error

Use outputs of best re-run

Perception model inaccuracy

Generate larger test-set

Re-train the model

There are no rows in this table

⁠

Order of investigation and why:

Control errors. This is performed first since it is the easiest to measure and if they are not present it is helpfult to check off first. Modelling errors such as bad kinematic models and incorrect joint taring can cause control errors, however it is better to investigate those with the context of whether the robot believes it is following it’s reference first.

If there are no control errors this is the next to be tested are perception errors, as regardless of subsequent errors an incorrect target will inevitably cause inaccuracy

Finally the kinematic and taring errors are investigated

Want to print your doc?
This is not the way.

Try clicking the ··· in the right corner or using a keyboard shortcut (

CtrlP

) instead.