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

Venki Parsuram; Michelle Wang; Jin-Hyuk Hong; Pulkit Agrawal

cs.ROOct 15, 2026

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

Authors: Venki Parsuram, Michelle Wang, Jin-Hyuk Hong, Pulkit Agrawal

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.

Figures & tables

Fig. 2: Overview of C-GMT training and deployment. During training, a privileged teacher generates compliant references from external force and stiffness commands, which are distilled into a proprioceptive residual adapter on top of a force-robust GMT. At deployment, the same adapter modifies the motion reference using only proprioceptive observations and the commanded stiffness.

Load	Tracker	Success	MPJPE-G	MPJPE-L	MPJPE-PA
No load	SONIC	0.953	135	28.8	20.9
	Robust	0.977	183	26.5	18.8
$2\,\mathrm{kg}$ , right hand	SONIC	0.953	154	29.9	22.9
	Robust	0.977	187	26.6	19.5
$4\,\mathrm{kg}$ , right hand	SONIC	0.914	208	33.0	25.6
	Robust	0.977	212	27.6	20.6

TABLE I: Comparison of tracking performance between SONIC and our force-robust motion tracker on 128 held-out motion clips under external loads. MPJPE values are reported in mm.

Fig. 3: C-GMT accurately realizes a range of stiffnesses. We apply external forces to the hands of a standing robot in simulation and measure the displacement. We plot the median-force displacement ratio for C-GMT and a standing SoftMimic policy across a range of stiffnesses. C-GMT and SoftMimic have comparable stiffness adherence.

Group	Method	Success	$\boldsymbol{\epsilon_{g,p}}$ [m]	$\boldsymbol{\epsilon_{g,r}}$ [rad]	$\boldsymbol{\epsilon_{l,p}}$ [m]	$\boldsymbol{\epsilon_{l,r}}$ [rad]
Walking	Ours ( $k=100$ )	97.1%	0.44 (0.36)	0.21 (0.08)	0.038 (0.014)	0.20 (0.07)
	Ours ( $k=300$ )	97.9%	0.43 (0.36)	0.21 (0.08)	0.037 (0.014)	0.20 (0.07)
	SONIC	98.9%	0.24 (0.22)	0.20 (0.06)	0.040 (0.010)	0.21 (0.05)
	Robust decoder (no adapter)	97.5%	0.37 (0.34)	0.19 (0.07)	0.032 (0.013)	0.18 (0.06)
	TWIST2	86.8%	1.00 (1.08)	0.45 (0.40)	0.030 (0.021)	0.19 (0.12)
	GentleHumanoid	68.6%	0.97 (0.74)	0.51 (0.38)	0.031 (0.022)	0.20 (0.12)

TABLE II: Tracking performance on LAFAN1 walking and dynamic motions without external force perturbations. Values in parentheses denote standard deviation.

Fig. 4: C-GMT reduces collision forces even during highly dynamic motions. We place the robot next to a wall and play a dancing motion clip. In (a), we plot the contact force of the robot’s hand on the wall over time and observe that C-GMT reduces the interaction force. In (b), we show post-impact snapshots and observe that C-GMT yielded to the contact, remaining in the same position and continuing to track the motion. Both SONIC and TWIST2 push themselves away from the wall.

Fig. 5: C-GMT safely yields to humans and the environment even while performing a highly dynamic dancing motion clip.

Fig. 6: Full-body compliant responses improve robustness. We exert an increasing downward force on the standing robot’s hand. C-GMT and SoftMimic squat down to fully comply with the force, while GentleHumanoid never bends its knees and eventually falls under a strong force.

Fig. 7: Compliance enables contact-rich teleoperation like opening a door, wiping a whiteboard, dragging a chair while walking, and holding a rubber ball. C-GMT is robust to inevitable operator error when contact dynamics are unknown and imperceivable to the operator.

Disturbance	SONIC	C-GMT ( $k=300$ )
No disturbance	48/50	49/50
$20\,\mathrm{N}$ pull, close to table	34/50	44/50
$40\,\mathrm{N}$ pull, close to table	30/50	41/50
$60\,\mathrm{N}$ pull, early in walk	28/50	46/50

TABLE III: VLA task success under external disturbances. Each condition is evaluated over 50 trials from the same initial states.

Fig. 8: Compliance improves robustness while executing high-level policy commands. While rolling out a VLA, an unexpected force is applied to the robot. C-GMT complies, recovers, and completes the task successfully, but SONIC is pulled off-course, stumbles forward, collides with the box, and fails to pick it up.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

IK Objective	Weight
Pushed wrist position	$5.0$
Foot position / orientation	$2.5$ / $0.5$
Center of mass ( $x,y,z$ )	$0.1$ / $0.1$ / $10^{-6}$
Pelvis orientation	$0.03$
Keypoint position / orientation	$0.01$ / $0.001$
Wrist joint regularization	$0.05$

Appendix

TABLE IV: Task weights used in the compliant teacher’s whole-body IK solver.

Parameter	Robust GMT	Adapter
Simulator	Isaac Lab	mjlab
Physics timestep	$5\,\mathrm{ms}$	$2\,\mathrm{ms}$
Policy frequency	$50\,\mathrm{Hz}$	$50\,\mathrm{Hz}$
Training algorithm / optimizer	PPO	Adam
Learning rate	$2\times 10^{-5}$ (actor)	$2\times 10^{-5}$
Environments	$49{,}152$	$32{,}768$

Appendix

TABLE V: Training settings for the force-robust tracker and residual adapter.

Loss term	Weight
Action imitation	$1.0$
Action imitation under force	$2.0$
Motion-token imitation	$0.25$
Auxiliary force prediction	$0.1$
Residual magnitude	$0.001$
No-force action regularization	$0.05$

Appendix

TABLE VI: Residual adapter loss weights. Additional rate and jerk penalties of

0.5

are applied on force-free samples.

Fig. 9: Effect of feasibility-based force sampling on compliance learning for a standing reference motion. Feasibility-based sampling achieves faster convergence and higher adherence to the commanded give-way compared with uniform force sampling.

Compliant Robot Control

See all 15 papers

PGMT: Perceptive General Motion Tracking for Humanoid Robots
Sep 8, 2026Hongyi Li, Li Peizhuo, Yucheng Tao +9Humanoid Robot Locomotion Terrain-Aware Robot Locomotion
Center of X-Mechanics, Zhejiang University, Hangzhou, China. · MARMot Lab, National University of Singapore, Singapore · MirrorMe Technology Co., Ltd.
What Matters in Humanoid General Motion Tracking? An Empirical Study
Jul 22, 2026Fabio Amadio, Enrico Mingo HoffmanHumanoid Robot Control Robotic Policy Evaluation
HUCEBOT · Inria, Université de Lorraine, CNRS, Nancy, France
OmniTrack: General Motion Tracking via Physics-Consistent Reference
Feb 27, 2026Yuhan Li, Peiyuan Zhi, Yunshen Wang +6Humanoid Motion Tracking Humanoid Robot Control
Huazhong University of Science and Technology · State Key Lab of General AI, Beijing Institute for General Artificial Intelligence (BIGAI) · Shanghai Jiao Tong University

Sep 8, 2026J/K move · Enter open · S save

PGMT: Perceptive General Motion Tracking for Humanoid Robots

Hongyi Li, Li Peizhuo, Yucheng Tao, Ze Wang+8

Center of X-Mechanics, Zhejiang University, Hangzhou, China. · MARMot Lab, National University of Singapore, Singapore · MirrorMe Technology Co., Ltd.

Humanoid Robot Locomotion Terrain-Aware Robot Locomotion Humanoid Motion Tracking

Open paper AskPDF arXiv

Humanoid motion trackers can reproduce diverse whole-body motions, but their performance degrades on complex terrain where terrain-agnostic references become physically infeasible. We present PGMT, a Perceptive General Motion Tracking pipeline for humanoid robots that learns terrain adaptation from independently selected motion references and terrains. PGMT first learns a general tracking and recovery prior, then incorporates terrain perception through motion-conditioned terrain glimpses that selectively encode regions relevant to the current motion. Terrain-aware tracking relaxation allows necessary deviations from the reference while preserving its motion intent. Zero-shot deployment on a Unitree G1 demonstrates robust terrain-adaptive locomotion and whole-body motion execution over real-world terrain with obstacles up to 37 cm high, while supporting teleoperation, dynamic motion tracking, and fall recovery. PGMT extends general humanoid motion tracking beyond flat ground, providing a unified policy for terrain-adaptive locomotion, diverse whole-body behaviors, and teleoperation in complex environments. Project homepage: https://luyili.github.io/pgmt/

Abstract

Compliant Robot Control

Explore similar work

PGMT: Perceptive General Motion Tracking for Humanoid Robots

What Matters in Humanoid General Motion Tracking? An Empirical Study

OmniTrack: General Motion Tracking via Physics-Consistent Reference