Abstract
Articulated object manipulation tasks such as opening doors, tightening bolts, and pulling drawers impose kinematic constraints on a robot end-effector's motion that neural policies may fail to satisfy. Unlike hybrid force-position control that strictly separates force and position control into orthogonal subspaces, impedance control modulates a mechanical impedance system in the form of a virtual inertia, damping, and stiffness. However, selecting task-appropriate impedance gains remains challenging for robot policies. In this paper, we address this problem for the case of manipulating articulated objects. Our constraint-decomposed impedance control method uses the sampling diversity of a generative policy to estimate the feasible motion directions in twist space. The key idea is to infer the feasible and constrained subspaces of twists from an ensemble of trajectories generated by the policy from the same observation. Principal component analysis (PCA) of the end-effector twists determines the feasible twist subspace, which is then used to form 6 × 6 impedance gain matrices.
For many common articulated objects, the feasible twist subspace is independent of the object configuration, so the subspace and resulting gains are estimated once and held fixed during manipulation, minimizing computational overhead. For less common cases where this independence does not hold, we instead estimate a configuration-independent twist subspace of higher dimension that contains the local feasible subspaces, allowing the same fixed-gain strategy to be used.
Experiments across simulated and real-world single- and multi-DoF objects show that the proposed method consistently reduces constraint-violating contact loads relative to isotropic impedance control while maintaining task execution.
Video
Method
When a policy's trajectory slightly violates the object's kinematic constraint, tracking it stiffly in every direction produces large contact and internal forces. The question is not only how stiff, but which directions should be stiff. We read the answer from the policy itself.
-
1
Trajectory ensemble
At grasp closure, the diffusion policy samples N trajectories from the same observation.
-
2
Twists
Consecutive poses of each sample give end-effector body twists, stacked into $W \in \mathbb{R}^{6\times NH}$.
-
3
Feasible subspace
Weighted, uncentered PCA of the twists gives the feasible twist subspace $U_\parallel$.
-
4
Decomposed gains
Impedance gains stay stiff along $U_\parallel$ and become compliant along its complement: $K_v = \bar\Lambda\,(k_\parallel P_\parallel + k_\perp P_\perp)$.
Revolute
Cylindrical
Planar
Universal
Revolute
Prismatic
Helical
Cylindrical
CrankSimulation: Step Through a Rollout
Four articulated objects in Isaac Sim, manipulated by a flying gripper. Every controller runs the same Diffusion Policy: Iso uses isotropic impedance, Ours lowers the impedance only along the estimated constrained complement, and Oracle uses the same decomposition with the analytic subspace (simulation only).
A representative trial per object (median Ours/Iso force-impulse ratio), replayed from the recorded states. Drag to orbit, scroll to zoom.
Ours relative to Iso constraint-violating load, mean over the trials in which all three controllers succeed
Iso (= 1.0)Ours, relative to Isoabove Iso Under each bar: mean load, Iso → Ours. The highlighted row follows the 3D view.
Ours lowers the force impulse to 0.13–0.37× and the peak force to 0.18–0.46× of Iso, recovering at least 90% of Oracle's reduction in both. A single multi-task policy trained on all four objects shows the same trend (0.14–0.37× and 0.20–0.49×).
Full simulation table
Triples are Iso / Ours / Oracle. SR counts all 100 trials; other entries are means over the trials in which all three controllers succeed. SE is reported for Ours. Dashes: the universal joint has no constrained rotation.
Real World
A Franka Research 3 manipulates five objects: four single-joint objects and a coffee grinder whose crank has a configuration-dependent feasible subspace.
Ours relative to Iso constraint-violating load, mean over 10 rollouts per controller
Iso (= 1.0)Ours, relative to Isoabove Iso Under each bar: mean load, Iso → Ours. † Prismatic has no feasible rotation, so its moments are the full moment. ‡ Prismatic excluded.
On the crank the improvement is smaller: the fixed envelope contains directions that are constrained at the current configuration, and the force impulse rises to 1.14× while the other three loads still decrease.
Full real-world table
Entries are Iso / Ours (Ours/Iso), averaged over the rollouts. SE is reported for Ours. † Prismatic has no feasible rotation, so ICM and PCM measure the full moment.
BibTeX
If you use this work, please cite the CoRL 2026 (PMLR) version.
@inproceedings{park2026constraint,
title = {Constraint-Decomposed Impedance Control from Generative Policy
for Articulated Object Manipulation},
author = {Park, Kyungseo and Lim, Byeongdo and Lim, Jungbin and
Choi, Minseok and Park, Frank C.},
booktitle = {Proceedings of The 10th Conference on Robot Learning},
series = {Proceedings of Machine Learning Research},
publisher = {PMLR},
year = {2026}
}