cs.ROAug 21, 2026

Demonstration-Guided Humanoid Stand-Up on an Emulated Deformable Surface

Authors: Aniruddh Kushwah, Vyankatesh Ashtekar, Ashish Dutta

Organizations: Department of Mechanical Engineering Indian Institute of Technology Kanpur Kanpur, Uttar Pradesh, India

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

This paper presents a reference-guided reinforcement learning framework to generate stand-up motion for a 29-DOF Unitree G1 humanoid on deformable soft ground, using a human demonstration recorded on hard ground. The terrain compliance is modelled using solref and solimp parameters from MuJoCo's rigid body soft-contact model. The rewards consists of (i) reference motion tracking through residual joint-position control and (ii) explicit recovery objectives such as pelvis height, torso uprightness, and the final posture. First, the policy is trained with the specified rewards considering hard ground. Next, the terrain stiffness is lowered by updating solref and the nominal surface penetration zone is expanded using solimp. Subsequent training enables the policy to adapt to the delayed support force generation due to significant surface penetration during contact-intensive phases while preserving the original demonstration pattern. The learned policy successfully completes the fallen-to-standing task in simulation, reaching the targeted pelvis height and uprightness, with a maximum contact penetration of approximately 40 mm during the process. The proposed method is demonstrated on two stand-up sequences and successfully achieves the final recovery objective on both hard and soft ground. Ablation studies show that reference tracking alone is insufficient for successful stand-up, and that explicit recovery rewards are essential.

Figures & tables

TermDefinitionWeight
Reference-Tracking Rewards
Joint pose trackingrtpose=exp⁡[−4mean⁡((qtsim−qtref)2)]\displaystyle r^{\mathrm{pose}}_{t}=\exp\left[-4\operatorname{mean}\left((\mathbf{q}^{\mathrm{sim}}_{t}-\mathbf{q}^{\mathrm{ref}}_{t})^{2}\right)\right]2.02.0
Joint velocity trackingrtvel=exp⁡[−0.03mean⁡((q˙tsim−q˙tref)2)]\displaystyle r^{\mathrm{vel}}_{t}=\exp\left[-0.03\operatorname{mean}\left((\dot{\mathbf{q}}^{\mathrm{sim}}_{t}-\dot{\mathbf{q}}^{\mathrm{ref}}_{t})^{2}\right)\right]0.70.7
Root planar trackingrtxy=exp⁡[−2mean⁡((ptxy−ptxy,ref)2)]\displaystyle r^{xy}_{t}=\exp\left[-2\operatorname{mean}\left((\mathbf{p}^{xy}_{t}-\mathbf{p}^{xy,\mathrm{ref}}_{t})^{2}\right)\right]0.20.2
Pelvis height trackingrth=exp⁡[−10(ht−htref)2]\displaystyle r^{h}_{t}=\exp\left[-10(h_{t}-h^{\mathrm{ref}}_{t})^{2}\right]1.01.0
Root orientation trackingrtori=exp⁡(−θt2)\displaystyle r^{\mathrm{ori}}_{t}=\exp(-\theta_{t}^{2})0.50.5
TABLE I: Reward terms used for reference tracking and final stand-up completion.
Fig. 1: Pelvis height and torso uprightness for one representative policy.
Fig. 2: Sequential snapshots of the stand-up policy in MuJoCo simulation environment.
Fig. 3: Active floor contacts for one representative policy: penetration depth and active-contact count. The maximum penetration was approximately 4040 mm.
MetricTracking onlyComplete reward
Aggregate RMS error [rad]0.15710.15710.15600.1560
Final pelvis height [m]0.0590.0590.792\mathbf{0.792}
Maximum pelvis height [m]0.1300.1300.794\mathbf{0.794}
Final uprightness0.5430.5430.991\mathbf{0.991}
Maximum uprightness0.9170.9171.000\mathbf{1.000}
TABLE II: Reward ablation for one representative trajectory

Explore similar work

CardsList

  1. Natural Sit-to-Stand Motion Synthesis For Humanoids via Guided Assistance Curricula and Staged Rewards

    Aug 21, 2026Meet Pal Singh, Vyankatesh Ashtekar, Ashish DuttaHumanoid Robot ControlHumanoid Robotics

    Department of Mechanical Engineering, Indian Institute of Technology Kanpur, Kanpur, Uttar Pradesh, India

  2. Terrain Consistent Reference-Guided RL for Humanoid Navigation Autonomy

    May 15, 2026William D. Compton, Zachary Olkin, Aaron D. AmesHumanoid Robot LocomotionReinforcement Learning

    Department of Computing and Mathematical Sciences, California Institute of Technology, Pasadena, CA 91125

  3. MILD: Tractable Terrain Modeling for Learning Improved Bipedal Locomotion on Deformable Surfaces

    Aug 20, 2026Zeren Luo, Jiahui Zhang, Zhe Xu +6Terrain-Aware Robot LocomotionHumanoid Robot Locomotion

    Adaptive Robotic Controls Lab (ArcLab), Department of Mechanical Engineering, The University of Hong Kong, Hong Kong · School of Mechanical and Electrical Engineering, Beijing Institute of Technology, Beijing, China · Graduate School of Information Science and Technology, The University of Tokyo, Tokyo, Japan