cs.ROSep 18, 2026

Learning Distance-Conditioned Object Transport for Humanoid Loco-Manipulation from a Single Motion Clip

Authors: Yuhyeon Hwang, Daniel Sungho Jung, YongHyeok Seo, Mingi Jung, Chang Nho Cho, Jung-Hoon Hwang, Dongin Shin

Organizations: Korea Electronics Technology Institute, Republic of Korea. · Seoul National University, Republic of Korea. · Korea University, Republic of Korea.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Motion tracking can reproduce humanoid loco-manipulation from a single retargeted motion clip, but a policy trained on a fixed reference primarily reproduces its demonstrated transport outcome. Although the source trajectory visits intermediate object displacements, transport termination is demonstrated only at its endpoint. We identify this mismatch as the termination-versus-passage gap: intermediate displacements are observed as passage states rather than termination-complete outcomes. We introduce Distance-Conditioned Reference Recomposition (DCRR), which relocates the demonstrated termination segment to intermediate transport states. A frozen tracking teacher replays the recomposed references under closed-loop dynamics, and the retained trajectories are relabeled by their achieved object placements and distilled into a reference-free policy. This procedure constructs distance-conditioned supervision from the interaction behavior encoded in the source motion. Across Carry, Kick-Push, Crouch-Push, and Drag, DCRR-BC produces command-dependent transport with an overall normalized distance mean absolute error (MAE) of 0.15, compared with 0.28 for source-only behavior cloning. RL fine-tuning further improves the command response and execution robustness in the training simulator and under sim-to-sim transfer. Finally, hardware experiments demonstrate transport-distance modulation across all four interaction modes.

Figures & tables

Fig. 1: The termination-versus-passage gap in a single source motion: intermediate displacements occur during transport, while termination is demonstrated only at the source endpoint.
Fig. 2: Four source motions for object transport: Carry, Kick-Push, Crouch-Push, and Drag.
Fig. 3: Three-stage DCRR pipeline: goal-conditioned teacher training; reference recomposition, closed-loop replay, filtering, and achieved-goal relabeling; reference-free policy learning through distillation and RL fine-tuning.
 dachd_{\mathrm{ach}}Valid replay (%)
dnomd_{\mathrm{nom}}Parallel++ Orth.Parallel++ Orth.
0.20.080.1228.690.7
0.40.200.2862.098.6
0.60.430.4894.699.4
0.80.710.7198.693.3
1.0 (Source)0.920.9294.794.8
TABLE I: Achieved distance and closed-loop replay validity of recomposed references, across four motions.
Fig. 4: Command-response curves across four interaction modes. Markers and shaded regions show the mean and one standard deviation, respectively; the diagonal denotes exact command-response correspondence ( d^=d\hat{d}=d ).
MethodCarryKick-PushCrouch-PushDragOverall
 MAE ↓\downarrowFall (%) ↓\downarrowMAE ↓\downarrowFall (%) ↓\downarrowMAE ↓\downarrowFall (%) ↓\downarrowMAE ↓\downarrowFall (%) ↓\downarrowMAE ↓\downarrowFall (%) ↓\downarrow
Source Teacher (Wide)0.46±0.260.46{\pm}0.260.030.030.33±0.230.33{\pm}0.230.070.070.56±0.390.56{\pm}0.390.00\mathbf{0.00}0.37±0.240.37{\pm}0.240.00\mathbf{0.00}0.430.430.030.03
Src-BC0.40±0.260.40{\pm}0.260.600.600.28±0.180.28{\pm}0.1829.7729.770.16±0.210.16{\pm}0.2112.7012.700.28±0.200.28{\pm}0.2014.3014.300.280.2814.3414.34
HS-BC0.54±0.290.54{\pm}0.290.030.030.42±0.280.42{\pm}0.285.735.730.41±0.280.41{\pm}0.281.271.270.38±0.220.38{\pm}0.2253.6353.630.440.4415.1715.17
AMP-Body+Object0.05±0.03\mathbf{0.05{\pm}0.03}0.300.300.04±0.04\mathbf{0.04{\pm}0.04}0.030.030.20±0.150.20{\pm}0.150.00\mathbf{0.00}0.53±0.290.53{\pm}0.290.00\mathbf{0.00}0.210.210.080.08
DCRR Teacher0.05±0.04\mathbf{0.05{\pm}0.04}0.00\mathbf{0.00}0.11±0.060.11{\pm}0.060.00\mathbf{0.00}0.12±0.07\mathbf{0.12{\pm}0.07}0.00\mathbf{0.00}0.13±0.05\mathbf{0.13{\pm}0.05}0.030.030.10\mathbf{0.10}0.01\mathbf{0.01}
TABLE II: Distance MAE and fall rate across four interaction modes. Overall values are macro-averaged across interaction modes.
 Goal label
Command ddNominalAchieved (ours)
0.20.09±0.020.09{\pm}0.020.10±0.030.10{\pm}0.03
0.40.16±0.060.16{\pm}0.060.13±0.070.13{\pm}0.07
0.60.19±0.070.19{\pm}0.070.20±0.100.20{\pm}0.10
0.80.25±0.060.25{\pm}0.060.21±0.110.21{\pm}0.11
1.00.18±0.060.18{\pm}0.060.18±0.040.18{\pm}0.04
TABLE III: Distance MAE under nominal and achieved goal labels. Both variants use identical replay trajectories and differ only in their goal labels.
 Training SimulatorSim-to-Sim
Methodd=0.2d=0.2d=0.4d=0.4d=0.6d=0.6d=0.8d=0.8d=1.0d=1.0d=1.2d=1.2Fall (%) ↓\downarrowd=0.2d=0.2d=0.4d=0.4d=0.6d=0.6d=0.8d=0.8d=1.0d=1.0d=1.2d=1.2Fall (%) ↓\downarrow
DCRR Teacher0.11±0.030.11{\pm}0.030.28±0.040.28{\pm}0.040.49±0.050.49{\pm}0.050.71±0.050.71{\pm}0.050.93±0.050.93{\pm}0.051.04±0.061.04{\pm}0.060.170.170.13±0.050.13{\pm}0.050.27±0.050.27{\pm}0.050.51±0.060.51{\pm}0.060.71±0.090.71{\pm}0.090.94±0.110.94{\pm}0.111.07±0.121.07{\pm}0.120.170.17
DCRR-BC0.15±0.030.15{\pm}0.030.31±0.090.31{\pm}0.090.50±0.110.50{\pm}0.110.73±0.160.73{\pm}0.160.90±0.180.90{\pm}0.181.08±0.231.08{\pm}0.238.838.830.13±0.070.13{\pm}0.070.21±0.110.21{\pm}0.110.41±0.180.41{\pm}0.180.60±0.230.60{\pm}0.230.80±0.280.80{\pm}0.280.96±0.320.96{\pm}0.3217.7817.78
DCRR-RLFT0.15±0.040.15{\pm}0.040.33±0.030.33{\pm}0.030.53±0.060.53{\pm}0.060.76±0.080.76{\pm}0.080.95±0.070.95{\pm}0.071.12±0.081.12{\pm}0.082.222.220.13±0.060.13{\pm}0.060.32±0.090.32{\pm}0.090.50±0.110.50{\pm}0.110.71±0.130.71{\pm}0.130.92±0.110.92{\pm}0.111.10±0.111.10{\pm}0.118.508.50
TABLE IV: Command response and fall rates in the training simulator and under sim-to-sim transfer. Reported d^\hat{d} values are mean ±\pm standard deviation over non-fall episodes. Values are averaged across four interaction modes.
 d=0.4d=0.4d=0.7d=0.7d=1.0d=1.0d=1.2d=1.2
Motiond^\hat{d}Falld^\hat{d}Falld^\hat{d}Falld^\hat{d}Fall
Carry0.36±0.060.36{\pm}0.060/50.67±0.050.67{\pm}0.050/50.96±0.070.96{\pm}0.071/51.09±0.081.09{\pm}0.081/5
Kick-Push0.22±0.130.22{\pm}0.130/50.54±0.160.54{\pm}0.161/50.72±0.170.72{\pm}0.171/50.86±0.160.86{\pm}0.162/5
Crouch-Push0.33±0.090.33{\pm}0.090/50.60±0.090.60{\pm}0.090/50.91±0.110.91{\pm}0.111/51.07±0.121.07{\pm}0.121/5
Drag0.29±0.140.29{\pm}0.140/50.60±0.130.60{\pm}0.130/51.01±0.091.01{\pm}0.090/51.16±0.101.16{\pm}0.101/5
TABLE V: Hardware command response of DCRR-RLFT across four interaction modes. Achieved distances are summarized as mean ±\pm standard deviation over non-fall trials.
Fig. 5: Hardware rollouts of DCRR-RLFT across interaction modes and transport commands.

Explore similar work

CardsList

  1. Humanoid-DART: Humanoid Loco-Manipulation using Diffusion-guided Augmentation through Relabeling and Tracking

    Jun 25, 2026Pranav Debbad, Kanish Thiagarajan, Victor Dhédin +2Humanoid Loco-ManipulationRobot Imitation Learning

    Munich Institute of Robotics and Machine Intelligence (MIRMI), Technical University of Munich (TUM), Germany.

  2. REFINE-DP: Diffusion Policy Fine-tuning for Humanoid Loco-manipulation via Reinforcement Learning

    Mar 14, 2026Zhaoyuan Gu, Yipu Chen, Zimeng Chai +11Humanoid Loco-ManipulationDiffusion Policy

    Institute for Robotics and Intelligent Machines, Georgia Institute of Technology, Atlanta, GA 30332 USA

  3. FetchMan: Learning Visual Humanoid Loco-Manipulation Policies from Simulated Experiences

    Aug 17, 2026Omar Rayyan, Zhi Li, Max Argus +4Humanoid Loco-ManipulationSim-to-Real Transfer

    University of California, Los Angeles · Allen Institute for AI · University of Washington