C-GMT Abstract Overview Paper Theme

Compliant GMT: Making General Motion Tracking Compliant via Residual Reference Adaptation

Venki Parsuram Michelle Wang Jin-Hyuk Hong Pulkit Agrawal

MIT CSAIL Improbable AI
Paper arXiv Code (soon)
C-GMT on a Unitree G1. One policy. Any motion. Safer to be around. All videos are shown at 1× speed.

Abstract

We propose Compliant GMT (C-GMT), a modular framework for full-body compliant responses in general motion tracking. As humanoid robots move into real world deployment, they not only need to perform any motion but also respond safely and robustly to interaction forces. Existing general motion tracking policies can imitate diverse motions but are trained to be stiff, leading to unexpected and unsafe behaviors under disturbance. Prior work in compliant humanoid control trades off compliance and generality. There exist examples of full-body compliant responses, but these are heavily curated to single-reference policies. On the other hand, general motion trackers (GMT) are trained to rigidly imitate massive motion capture datasets. Our key insight is to leverage the generality of pretrained GMTs for motion tracking while introducing compliance from a residual adapter that conditions on a history of proprioception. This enables our approach, C-GMT, to achieve both general motion tracking and whole-body compliant responses. Compared to state-of-the-art general motion trackers and single-motion compliant policies, C-GMT demonstrates competitive performance in both motion tracking accuracy and compliance adherence. We further demonstrate improved safety, robustness, and error tolerance in simulation and real-world G1 experiments.

Overview

The problemGeneral motion trackers are trained to be stiff

General motion trackers are trained to closely follow reference motions, even under external disturbances, making them stiff. While this enables accurate motion tracking, it can also cause robots to resist unexpected contact, exerting large forces instead of yielding.

At high stiffness, the robot resists contact rather than yielding, knocking over the tower.
Held mid-dance, the stiff policy fights the hand instead of giving way.
Prior workExisting compliant controllers trade off generality and compliance

Existing approaches to compliant humanoid control face a tradeoff between generality and compliance. Methods that achieve whole-body compliant responses are limited to individual motions or simplified motion commands, while general motion trackers with compliance restrict their responses to the upper body.

Our approachFormulate compliance as a reference adaptation problem

We distill the compliance into a residual adapter, preserving the generality of a pretrained motion tracker while enabling whole-body compliant responses.

C-GMT method diagram. Top row: a privileged teacher takes a reference motion through a compliance IK solve, using the true force and commanded stiffness, to produce a compliant reference, then through a frozen force-robust GMT to the teacher action. Bottom row: the student feeds reference motion and proprioception into a residual adapter, sums the residual into the reference, and passes it to the same frozen tracker, with the adapter trained to match the teacher's action.
Teacher above, student below. The teacher uses the true contact force to compute a compliant reference. The student learns the same correction from proprioception alone and passes it to the same frozen tracker.
ResultTrack any motion

Compliance without sacrificing generality. C-GMT tracks diverse, dynamic whole-body motions, from teleoperation to dance.

Whole-body teleoperation.
Salsa, from a reference motion.
ResultBe safe while being dynamic

Even during fast, dynamic motion, C-GMT gives way when someone holds or pulls on it. A stiff tracker on the same motion fights back.

C-GMT (ours)
Stiff
ResultC-GMT allows accurate, adjustable compliance

Different tasks require different levels of compliance. C-GMT enables adjustable robot stiffness while accurately achieving the desired response.

k = 60
Low stiffness: the arm yields to the tower and leaves it standing.
k = 600
High stiffness: the same policy holds position and pushes through.
Plot of measured stiffness against commanded stiffness, showing the measured response following the commanded value across the range.
Measured vs. commanded stiffness for a standing motion. C-GMT, a single general motion-tracking policy, achieves stiffness adherence comparable to SoftMimic, trained specifically for the standing motion. Dashed line: ideal adherence.
ResultC-GMT improves safety and robustness

C-GMT improves safety and robustness, allowing the robot to yield to unexpected pushes and pulls and continue performing its task.

ResultCompliance helps manipulation

Compliance can also improve manipulation, making robots more robust to controller errors and enabling contact-rich tasks like opening doors and inserting objects.

Pushing in a misaligned tray.
Compliant search and insert in a movable block.
Opening articulated objects.
ResultC-GMT enables natural and fun physical interactions

C-GMT makes physical interaction more natural, whether you're moving the robot out of the way or simply having fun interacting with it.

Moving the robot out of the way while it wipes the counter.

Push it, pull it, play with it.

Paper

Compliant GMT: Making General Motion Tracking Compliant via Residual Reference Adaptation
Venki Parsuram, Michelle Wang, Jin-Hyuk Hong, Pulkit Agrawal
arXiv:2610.22119 · PDF

@article{parsuram2026compliant,
  title   = {Compliant GMT: Making General Motion Tracking
             Compliant via Residual Reference Adaptation},
  author  = {Parsuram, Venki and Wang, Michelle and Hong, Jin-Hyuk and Agrawal, Pulkit},
  journal = {arXiv preprint arXiv:2610.22119},
  year    = {2026}
}

Acknowledgments

This research was financially supported by the Ministry of Trade, Industry, and Energy (MOTIE), Korea, under the “Global Industrial Technology Cooperation Center program” supervised by the Korea Institute for Advancement of Technology (KIAT) (Grant No. P0028435).