Learn Weightlessness: Imitate Non-Self-Stabilizing Motions on Humanoid Robot

Yucheng Xin, Jiacheng Bao, Haoran Yang, Wenqiang Que, Dong Wang, Junbo Tan, Xueqian Wang, Bin Zhao, Xuelong LiarXiv 2026 · arXiv

TL;DR

Humanoid robots excel at balancing while walking, but struggle with tasks like sitting or leaning where they must rely on the environment for stability. This paper introduces a "Weightlessness Mechanism" (WM) that mimics human behavior by selectively relaxing specific robot joints, allowing the robot to safely "free-fall" into chairs or against walls. This approach enables a single whole-body control policy to generalize across various furniture heights and angles without any task-specific reward tuning.

What Problem Does This Paper Address?

Humanoid robot motions can be broadly split into two categories:

  1. Self-Stabilizing (SS): Walking, turning, or dancing, where the robot maintains balance entirely on its own.
  2. Non-Self-Stabilizing (NSS): Sitting, lying down, or leaning against a wall, where the robot must interact with and rely on external objects to stay upright.

While current imitation learning (IL) and reinforcement learning (RL) methods handle SS motions well, they fail miserably at NSS motions. Naive imitation learning forces rigid trajectory tracking, resulting in "fake-sitting" or "fake-leaning"—where the robot visually mimics the pose but fails to make actual, weight-bearing contact with the surface. If the chair's height or the bed's angle changes even slightly from the training data, the robot falls over. Conversely, RL requires tedious, task-specific reward engineering for every single scenario.

This paper bridges the gap by introducing a biologically inspired mechanism: allowing specific joints to enter a "weightless" (relaxed) state to exploit passive environmental contact for stability.

Methods

The authors developed a two-stage framework to implement this weightlessness mechanism:

  1. Weightlessness-State Auto-Labeling: To avoid manual annotation, the system automatically detects when and where a robot should relax. It identifies the weightless interval by checking when the robot's Center of Mass (CoM) projection falls outside its foot support polygon while making external contact. It then uses a limb-node hierarchy to determine which joints to relax. For example, if a robot leans on its left elbow, the joints from that elbow to the waist remain active to provide support, while the remaining limbs are flagged to relax.
  2. WM Network Training: An LSTM-based network is trained on this auto-labeled data. It takes the robot's pose history, current state, and future motion trends to predict a continuous relaxation vector ($w_t \in [0,1]$) for each joint, where $0$ means fully relaxed and $1$ means fully active. A temporal smoothness loss is added to prevent jerky transitions.
  3. Two-Stage Policy Fine-Tuning: First, a base whole-body imitation policy is trained using asymmetric Proximal Policy Optimization (PPO). Second, this policy is fine-tuned by integrating the WM network, directly scaling the low-level PD controller's output torque by the predicted relaxation vector ($\tau_t = \tau(a_t) \cdot w_t$).
[Base Imitation Policy (PPO)] -> [Fine-Tuning with WM Network (LSTM)] -> [Torque Modulation (τ * w)]

Datasets

The authors collected a custom dataset consisting of 200 video demonstrations split equally across four non-self-stabilizing action categories (50 videos each):

  • Sitting on chairs
  • Lying down on beds
  • Leaning against walls via the shoulder
  • Leaning against walls via the elbow

Human motions (SMPL format) and 3D scene geometries were extracted using off-the-shelf vision frameworks (GVHMR, MegaSaM, NKSR, VideoMimic) and retargeted to the target robot skeleton.

Key Insights and Ideas

  • Controlled Free-Fall is a Feature, Not a Bug: Instead of fighting gravity to rigidly follow a trajectory, robots can leverage gravity. By relaxing the lower limbs during a sit, the upper body naturally and safely "free-falls" onto the chair to complete the motion.
  • Fine-Tuning Beats Training from Scratch: The authors discovered that introducing joint relaxation from the very beginning of training hurts performance. Early random exploration triggers premature weightless states, leading to overly conservative behaviors. Training a rigid policy first and then fine-tuning it with the WM network yields significantly better stability.
  • Kinematic Tree Relaxation: Using the robot's kinematic tree to trace paths from the contact point to the waist provides an elegant mathematical way to isolate structural support joints from joints that can safely go limp.

Key Results

The framework was validated in simulation (Isaac Gym and MuJoCo) and successfully deployed on a real Unitree G1 humanoid robot.

  • Sitting Task: Generalizing across chair heights from 0.2m to 0.3m, the baseline imitation policy's success rate dropped to a dismal 8% at certain heights. The proposed WM framework maintained a robust 84% success rate.
  • Lying Down Task: Tested across bed inclinations of 0°, 45°, and 90°, the WM method achieved a 90% success rate, compared to 64% for the baseline.
  • Leaning Task: The robot successfully leaned against walls using its shoulder (82% success rate) or elbow (86% success rate) while actively resisting external perturbations applied to its non-supporting leg.

Limitations and Future Work

While the paper demonstrates excellent sim-to-real transfer using domain randomization, a few limitations are apparent:

  • Perception Reliance: The system heavily relies on accurate external scene perception (via the Unitree G1's built-in LiDAR) to construct heightmaps. Poor lighting or complex geometries could degrade the alignment.
  • Scope of Actions: The framework was evaluated on a discrete set of common human behaviors (sitting, lying, leaning). Future work will need to explore how this mechanism scales to highly dynamic, unpredictable contact scenarios, such as stumbling, catching a falling object, or physical collaboration with humans.

Summary updated

Abstract

The integration of imitation and reinforcement learning has enabled remarkable advances in humanoid whole-body control, facilitating diverse human-like behaviors. However, research on environment-dependent motions remains limited. Existing methods typically enforce rigid trajectory tracking while neglecting physical interactions with the environment. We observe that humans naturally exploit a "weightless" state during non-self-stabilizing (NSS) motions--selectively relaxing specific joints to allow passive body--environment contact, thereby stabilizing the body and completing the motion. Inspired by this biological mechanism, we design a weightlessness-state auto-labeling strategy for dataset annotation; and we propose the Weightlessness Mechanism (WM), a method that dynamically determines which joints to relax and to what level, together enabling effective environmental interaction while executing target motions. We evaluate our approach on 3 representative NSS tasks: sitting on chairs of varying heights, lying down on beds with different inclinations, and leaning against walls via shoulder or elbow. Extensive experiments in simulation and on the Unitree G1 robot demonstrate that our WM method, trained on single-action demonstrations without any task-specific tuning, achieves strong generalization across diverse environmental configurations while maintaining motion stability. Our work bridges the gap between precise trajectory tracking and adaptive environmental interaction, offering a biologically-inspired solution for contact-rich humanoid control.
BibTeX
@misc{xin2026learn,
  author = {Yucheng Xin and Jiacheng Bao and Haoran Yang and Wenqiang Que and Dong Wang and Junbo Tan and Xueqian Wang and Bin Zhao and Xuelong Li},
  title = {{Learn Weightlessness: Imitate Non-Self-Stabilizing Motions on Humanoid Robot}},
  year = {2026},
  eprint = {2604.21351},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.21351},
}

Loading related content...

Figure