Home People Research Publications Join
← Back to Research

Robotic manipulation · Iterative learning control

Differentiable Iterative Learning Control

for Black-Box Position-Commanded Manipulators

Diff-ILC learns how to improve a robot's commanded trajectory from repeated executions, without access to its low-level servo dynamics. A constrained ILC update keeps each correction feasible, while differentiating through that update adapts the learning controller between physical trials.

Mohammed Al-Ghaili · Yakoop Qasim · Hassan Almubarak

Control and Instrumentation Engineering Department · KFUPM · Dhahran, Saudi Arabia

Physical tracking experiments on the DOBOT CR3 across the four primary evaluation trajectories.
78.8% Lower mean joint-vector RMSE
76.4% Lower mean Cartesian RMSE
2 updates to achieve near-millimeter tracking accuracy

Trial-3 results for Diff-ILC with ten outer optimization epochs, relative to fixed-parameter constrained model-based ILC; averages over the four primary 0.1 Hz trajectories.

Video

Three-minute project overview

A compact walkthrough of the problem, the differentiable constrained-ILC loop, and the physical robot experiments.

Project overview: why repetitive tracking is difficult on black-box position-commanded manipulators, how Diff-ILC adapts its constrained learning update, and what changes on the physical robot.
Idea

Learning from each repetition

Commercial manipulators often expose joint-position commands while keeping their embedded servo controllers and actuator-level dynamics inaccessible. That makes it difficult to improve tracking by redesigning the low-level controller directly.

Diff-ILC instead learns at the command level. After each physical execution, it uses the measured tracking error to compute the next feasible command correction, then adapts the ILC design parameters computationally before the robot moves again.

Black-box command interface

Uses commanded joint positions and measured robot motion; no torque-level model or access to the embedded servo controller is required.

Feasible updates

The ILC correction respects joint-position, command-velocity, acceleration, and trial-to-trial correction limits.

Learning between motions

Several parameter-update epochs can run from the same measured trial, without requiring extra physical robot executions.

Graphical summary of the Diff-ILC forward execution, backward parameter update, robot tracking task, parameter tuning, and physical-trial convergence
Diff-ILC at a glance: a constrained ILC update produces the next command, while a task-level loss and KKT hypergradient adapt the ILC parameters between physical trials.
Method

How Diff-ILC closes the loop

The method separates expensive physical executions from computational adaptation: the robot supplies the measured trial, and the learning controller can update several times before the next motion.

1

Identify once

Paired smooth perturbations identify a compact differential model of how command changes affect measured motion. The model is then kept fixed during the ILC experiments.

2

Execute and correct

The current command is executed on the robot. From the measured error, a finite-horizon constrained QP computes a feasible trial-to-trial correction.

3

Adapt through the optimizer

A real-anchored prediction evaluates the candidate next trial in joint space, Cartesian space, and command smoothness. Hypergradients are obtained by differentiating the locally active KKT system.

4

Move only after learning

After the computational parameter updates, the constrained ILC problem is re-solved with the adapted parameters. Only that final command is executed on the next physical trial.

Trial-level computational graph. The measured robot trajectory anchors the next-trial prediction; the backward pass updates the ILC design parameters without another robot execution. Click to view full resolution.
What is adapted? Diff-ILC learns 19 ILC design parameters: six correction penalties, six first-difference penalties, six second-difference penalties, and one shared trial-to-trial correction bound.
Experiments

One identified model, several motions

The evaluation uses a 6-DoF DOBOT CR3 with only its joint-position command interface. The same identified differential model is reused across unseen planar and three-dimensional trajectories, and again for a faster execution condition.

Circle, infinity, star, and 3D helix tracking using the same identified differential model; no trajectory-specific re-identification is performed.
Robot
6-DoF DOBOT CR3
Interface
Joint-position commands
Control rate
100 Hz
Primary motions
0.1 Hz · ≈20 s
Physical ILC trials
10 per experiment
Faster-motion test
0.15 Hz · 1.5× faster

The differential model is identified once using three nominal executions and six symmetric perturbation directions (15 physical identification trials total), then kept fixed throughout the ILC evaluations.

Results

Faster improvement in the first few physical trials

Across the four primary trajectories, the ten-epoch Diff-ILC configuration reaches a mean joint-vector RMSE of 0.35° and Cartesian RMSE of 1.06 mm at Trial 3. The fixed-parameter baseline reaches 1.65° and 4.50 mm, respectively.

Joint-vector and Cartesian tracking RMSE over the first five physical trials for circle, infinity, star, and helix trajectories
Tracking convergence over the first five physical trials. The main advantage of Diff-ILC appears early: it reduces tracking error much faster than the fixed-parameter baseline.
Physical-trial efficiency: by Trial 3—after only two trial-to-trial command updates—Diff-ILC already achieves lower mean tracking error than the fixed-parameter baseline reaches by Trial 5.
Trial 3: Cartesian position RMSE
Trajectory Fixed MB-ILC Diff-ILC (10 epochs) Reduction
Circle4.05 mm0.73 mm82.0%
Helix3.79 mm0.73 mm80.7%
Infinity4.48 mm0.95 mm78.8%
Star5.68 mm1.82 mm68.0%
Fast circle (0.15 Hz)6.02 mm1.45 mm75.9%

Values are means over three repeated experiments. The fast-circle condition is not included in the four-trajectory averages used for the headline metrics.

A harder path: the star trajectory

The star path introduces sharp changes in direction. At Trial 3, Diff-ILC reduces joint RMSE from 2.06° to 0.52° and Cartesian RMSE from 5.68 mm to 1.82 mm relative to fixed-parameter model-based ILC.

Joint-wise tracking-error time histories on the star trajectory at Trial 3 comparing fixed model-based ILC and Diff-ILC
Trial-3 joint-wise tracking errors on the star trajectory.

How the learned parameters evolve

On the star trajectory, the correction penalties decrease while temporal-variation and curvature penalties increase. The shared correction bound grows from 1° to 8.29°, allowing larger corrections while increasingly penalizing nonsmooth command variation.

Evolution of Diff-ILC correction, temporal-variation, curvature penalties, and the shared correction bound
Online evolution of the learned Diff-ILC parameters for the star trajectory.
Robot trials

Trajectory tracking experiments

The videos below show the physical DOBOT CR3 experiments, including the baseline and Diff-ILC tracking evolution across repeated executions.

Primary trajectories

Four unseen evaluation paths executed at 0.1 Hz using the same identified differential model.

Circle

Planar circular reference trajectory · 0.1 Hz.

Infinity

Planar infinity-shaped reference trajectory · 0.1 Hz.

Star

High-curvature star-shaped reference trajectory · 0.1 Hz.

3D helix

Three-dimensional helical reference trajectory · 0.1 Hz.

Faster-motion transfer

The circle is then executed at 0.15 Hz (1.5× faster) using the same identified differential model, without speed-specific re-identification. At Trial 3, Diff-ILC reduces Cartesian RMSE from 6.02 mm to 1.45 mm.

Fast circle

Transfer to a higher execution rate with the same fixed differential model.