Bipedal Balance Control with Whole-body Musculoskeletal Standing and Falling Simulations

Chengtian Ma,   Yunyue Wei,   Chenhui Zuo,   Chen Zhang,   Yanan Sui
Tsinghua University     
Conference on Robot Learning (CoRL), 2025
Paper Code (Comming soon)

Bipedal balance is actually harder than it looks.

For bipedal robots, it's hard to control.
For human-beings, it takes time to master.

We studied the difficulty of human balance through a realistic musculoskeletal model, providing:
• A high-dimensional control method for stable bipedal balance
• Fine-grained measurements and analysis of adaptive balance strategies
• Large-scale simulation of different human balance and falling behaviors

We used MS-Human-700, a comprehensive whole-body model consisting of 90 rigid body segments, 206 joints, and 700 muscle-tendon units.

High-dimensional Musculoskeletal Balance Control Difficulties

The predominant approaches to musculoskeletal control are deep reinforcement learning (DRL). However, we found that:

Intuitive reward design is difficult, and natural standing behavior is not learned after training.

Long training time limits the possiblity of large-scale data collection.

Failure of DRL-based method on balance control

The Hierarchical Balance Control (HBC) Method

We propose an effective method for human musculoskeletal balance control.

The HBC Method: Our approach separates control into two hierarchical levels. The high-level planner provides control targets, while the low-level controller converts these targets into muscle control commands, directing the musculoskeletal model to achieve the desired balance.

Validation: Our simulated muscle activation patterns align well with real-world human experiments, particularly in calf muscle EMG measurements.

Key Advantages: HBC achieves stable, natural human musculoskeletal balance control with near real-time performance. The method is training-free and suitable for large-scale data collection of balance behaviors.

HBC Algorithm

Fine-grained Measurements and Analysis

We collected simulated balance and fall with 100 times scale improvement from real-world datasets, and made it possible for multi-modal balance and fall measurements.

Data Collection: The collected data includes fine-grained center of mass trajectories and muscle dynamics.

Measurement 1

Balance Region Analysis: With the center of mass data, we measured a balance region formed by center of mass trajectories, which will converge within the polygon of support.

Measurement 2
Measurement 3

Clinical Validation: We also calculated the distribution of collision position of fall, which matches with clinical data.

Measurement 4

Multi-scenario Simulation of Human Balance and Falling Behaviors

To extend the simulation to more scenarios, we collected data with injured models, and found a common shrinkage of the balance region.

Rectus femoris injury

Peroneus longus injury

Adductor magnus injury

Measurement 4
Measurement 4
Measurement 4


We deployed real-time hip exoskeleton assistance, successfully improved the balancing performance under perturbation, and reduced muscle effort.

Measurement 4

Exoskeleton assistance improves balance performance and reduces muscle activation levels.

BibTeX

@misc{ma2025bipedalbalancecontrolwholebody,
          title={Bipedal Balance Control with Whole-body Musculoskeletal Standing and Falling Simulations}, 
          author={Chengtian Ma and Yunyue Wei and Chenhui Zuo and Chen Zhang and Yanan Sui},
          year={2025},
          eprint={2506.09383},
          archivePrefix={arXiv},
          primaryClass={cs.RO},
          url={https://arxiv.org/abs/2506.09383}, }