Learning Deformable Object Manipulation Using Task-Level Iterative Learning Control

Venue: RSS 2026 (Sydney, Jul 13–17) · Session: Manipulation 3 · paper #126 Authors: Krishna Suresh, Christopher G. Atkeson arXiv: 2602.21302 · program page

Summary compiled from the arXiv paper (v2, posted under the earlier title "Learning Dynamic Rope Manipulation Using Task-Level Iterative Learning Control" — same authors and content); all numbers quoted from the paper. Trend context: RSS 2026 survey.

Stages of a flying knot tied by a human (top) and an xArm 7 robot arm (bottom) over 0.56 s: the hand/gripper moves up and twists to form a loop, then arcs so the weighted rope end flips through the loop, tying an overhand knot in a single one-handed motion.

Problem

Dynamic manipulation of deformable objects like ropes is hard because they have many unactuated degrees of freedom and are expensive to model; classical model-based ILC, which weights trajectory-tracking errors equally, fails on such tasks. The case-study task is the "flying knot": tying an overhand knot in mid-air with one continuous arm motion.

Method

Task-Level ILC (CMU) learns directly on hardware from a single human demonstration and a deliberately simplified point-mass rope model (maximal-coordinate variational integrator; one fixed parameter set for all ropes). Two key ideas: (1) a critical-point objective — instead of weighting errors along the whole trajectory, learning targets the rope state at one critical moment (the rope's self-collision that forms the loop); and (2) object trajectory learning — corrections are propagated to the unactuated rope state, not just the robot trajectory. Each iteration, a QP inverse model (built in Drake) maps the measured critical-point error to a spline-knot command correction subject to joint position/velocity/acceleration/torque limits. Hardware: xArm 7 with Vicon Vantage 16 motion capture at 200 Hz.

Results

Across 7 rope types — chain, latex surgical tubing, braided and twisted ropes, 7–25 mm thick, 0.013–0.5 kg/m — learning achieves a 100% success rate within 10 trials on all ropes from the same single demonstration, and a learned command replays at 100% success over 40 repeat trials. Transfer of a learned command between most rope pairs takes roughly 2–5 trials (many transfer in 0–2; two hard pairs exceeded 10). The equally-weighted ILC objective ablation fails the task, and learning also succeeds across 4 demonstration variants (Fast/Slow/Swipe/Inverse, 0.69–1.04 s).

Significance

A counterpoint to data-hungry learned policies: one demonstration, a crude physics model, and ~10 hardware trials suffice for a genuinely dynamic deformable-object skill — with the insight that where in the trajectory you place the learning objective matters more than tracking fidelity. Complements the deformable/contact-rich threads in Review-Dexterous-Manipulation.

← Back to RSS 2026 survey · RSS-2026-Papers · Home