Human-Like Redundancy Resolution

for Real-Time Bimanual Humanoid Teleoperation

Chao Tan,
Ming Shin Cindy Sung,
Shuxian Liang†,
Chaoqun Wan,
Zhen Huang,
Xian-Sheng Hua†

Institute of AI for Engineering, Tongji University † Corresponding authors.

TL; DR

We enable natural and stable bimanual humanoid teleoperation by combining adaptive human-like elbow guidance with singularity-robust wrist tracking.

Abstract

Real-time bimanual humanoid teleoperation requires accurate wrist tracking and stable arm posture. Kinematic redundancy allows different elbow configurations for the same wrist pose. Existing methods that track only the wrists can therefore cause elbow drift or configuration jumps. Anthropomorphic constraints help guide elbow posture, but fixed constraints can restrict singularity avoidance and impose unreliable guidance as the human arm straightens. We propose a human-like redundancy-resolution framework that prioritizes singularity-robust wrist tracking and uses adaptive elbow guidance to resolve redundancy. Shoulder–wrist geometric retargeting generates robot-scale wrist targets. Elbow-offset direction guides posture, and its magnitude controls guidance strength. Guidance weakens and deactivates as the human arm straightens. A hierarchical velocity solver then applies adaptive elbow guidance in the wrist-task null space. In a fixed-wrist-target ablation, adaptive guidance reduces mean elbow-direction error by 54.4%. In a workspace-boundary stress test, singularity-robust tracking reduces peak raw joint-velocity demand by 97.9%. Real-world bimanual tasks and cross-platform retargeting demonstrate practical utility across robot morphologies.

01

Why Redundancy Resolution Matters

A 7-DoF humanoid arm can reach the same 6-DoF wrist pose with many different elbow configurations. Wrist tracking alone therefore leaves the elbow underconstrained, which can cause unnatural drift or sudden configuration changes.

Adding a fixed anthropomorphic constraint does not fully solve the problem. Near full arm extension, the human elbow direction becomes unreliable, while rigid posture guidance can prevent the robot from using its redundancy to avoid singularities and joint limits. The controller must preserve the operator's wrist intent, recover a human-like posture when the observation is reliable, and release that guidance when it is not.

Comparison of wrist-only inverse kinematics, fixed anthropomorphic constraints, and the proposed human-like redundancy-resolution method
Wrist-only tracking underconstrains the elbow, while a fixed anthropomorphic constraint can restrict feasible wrist motion.

02

Our Method

Our human-like redundancy-resolution framework translates human motion to robot-scale wrist targets, adapts elbow guidance to the reliability of human posture, and preserves wrist-task priority through a hierarchical solver.

Overview of shoulder-wrist geometric retargeting, adaptive elbow-offset guidance, singularity-robust wrist tracking, and the hierarchical solver
Framework overview: human motion input, adaptive redundancy guidance, and real-time robot execution.
1

Geometric Retargeting

Shoulder–wrist geometry maps human upper-arm motion and wrist orientation into robot-scale targets while retaining the operator's end-effector intent.

2

Adaptive Elbow Guidance

Elbow-offset direction guides anthropomorphic posture. Its influence progressively weakens and deactivates as the human arm approaches full extension.

3

Robust Wrist Tracking

A hierarchical velocity solver suppresses near-singular velocity amplification and applies posture guidance only within the wrist task's null space.

03

Quantitative Evaluation

We evaluate 15 static arm poses, dynamic motion sequences, fixed-wrist ablations, and a near-boundary singularity stress test.

54.4%lower mean elbow-direction error
97.9%lower peak raw joint-velocity demand
2.87 mmmean static wrist position error
2.42°mean static arm-plane error
Static position, rotation, arm-plane error, and joint margin plots
Static benchmark. Tracking accuracy and arm-plane consistency across 15 poses.
Dynamic position and rotation error, arm-plane error, and singularity margin comparison
Dynamic benchmark. Stable wrist tracking and posture behavior throughout motion.
Elbow-direction error ablation plots
Elbow-guidance ablation. Adaptive guidance improves human–robot posture consistency.
Joint-velocity comparison near a singularity
Singularity ablation. Robust tracking prevents extreme joint-velocity demands.

04

Experiments

Simulation-Based Validation

Simulation validates bimanual wrist tracking, human-like redundancy resolution, and stable joint motion before real-robot deployment.

Real-Robot Bimanual Tasks

Wide-range motion

Stable across the workspace

From neutral posture through full extension and near-singular configurations, the controller maintains smooth motion and recovers naturally.

Coordinated manipulation

Ring stacking

Both arms coordinate grasping, transport, stacking, and placement while preserving stable elbow configurations.

Tool use

Cucumber peeling

One arm stabilizes the object while the other executes a constrained tool trajectory, requiring accurate bimanual coordination.

05

BibTeX

@misc{tan2026humanlike,
  title         = {{Human-Like Redundancy Resolution for Real-Time Bimanual Humanoid Teleoperation}},
  author        = {Tan, Chao and Sung, Ming Shin Cindy and Liang, Shuxian and Wan, Chaoqun and Huang, Zhen and Hua, Xian-Sheng},
  year          = {2026},
  eprint        = {2610.16855},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.16855}
}