Interactive research demo

Data-Driven Shape Control

Estimate how robot motion changes shape, then use that local model to drive the object toward a target.

Explore first

What should count as the same shape?

Predict what each representation will notice, then use Translate, Rotate and Bend to test your idea.

Start with TranslatePoint coordinates should change. Edge vectors and curvature should stay nearly unchanged.

An offline identification phase supplies the local input–shape relation; the online phase repeatedly measures error and corrects it.

Offline phase1. Probe the objectApply small robot commands and observe the resulting shape change.Δu → Δs

Offline phase2. Estimate the JacobianFit the local mapping from robot motion to feature motion.Ĵ ≈ Δs / Δu

Online phase3. Measure shape errorCompare the current and target feature vectors.e = s* − s

Online phase4. Correct and repeatConvert that error into a robot command, observe again and update.u̇ = λ Ĵ† e

Drag a gripper, its ring handle, or the object itself.

Object Target shape Measured features Discrete Kirchhoff rod · 24 segments

Shape match0%

Moves0

Move one gripper at a timeDrag a circular gripper point to translate it, or its small ring handle to rotate it.

Compare each signal with itself before and after a movement; the three rows use different units and scales.

‖e‖0.000

Rank—

Condition—

Idle — press Play to probe and servo.

6 × 18

No model yet. Probing applies a small motion on each of the six commands and measures the feature response.

Each feature component keeps one scale across the full time window; the error panel compares all three representations on the same normalised axis.

What changes between representations?

Keep position, orientation and shape. Best for placement, but any rigid motion shows up as error.

eₚ = p* − p

Remove translation while keeping orientation and local direction changes. Translating the object leaves this error untouched.

e_d = d* − d

Remove translation and rotation, isolating intrinsic bending. Shape converges, but pose becomes unobservable and uncontrolled.

eκ = κ* − κ

Takeaway The representation decides what the robot tries to correct.

Points preserve pose and shape, edges ignore translation, and curvature isolates bending. The “best” representation depends on the task.