Lingbao CASBOT and CASIA jointly propose the method of "enhancing human intent estimation and human-machine role allocation in physical human-machine collaboration"

Recently, the global conference ICRA 2025 in the field of robotics and automation opened in Atlanta, USA. The method of "Enhancing Human Intent Estimation and Role Allocation for Physical Human Robot Collaboration" jointly proposed by the Institute of Automation of the Chinese Academy of Sciences and Lingbao CASBOT was accepted by ICRA 2025.

01. DTRT: Enhancing Human Intent Estimation and Human Machine Role Allocation in Physical Human Machine Collaboration

Accurate human intent estimation and reasonable human-machine role allocation are crucial for safe and efficient collaboration in physical human robot collaboration (pHRC). The existing methods that rely on short-term motion data for intention estimation lack multi-step prediction ability, which hinders the ability of robots to perceive long-term intention changes and autonomously adjust human-machine allocation, leading to potential human-machine divergence. In order to solve these problems, the Institute of Automation of the Chinese Academy of Sciences and Lingbao CASBOT jointly proposed a dual transformer based robot tracker (DTRT) framework. The framework adopts a hierarchical structure, and uses human guided motion and force data to quickly capture human intention changes, so as to achieve accurate trajectory prediction and dynamic robot behavior adjustment, and ultimately achieve effective physical human-computer cooperation.

Robot tracker based on dual transformers for human intent estimation and role assignment

Specifically, human intention estimation in DTRT utilizes two Transformer based Conditional Variational Autoencoders (CVAE) to combine the robot's motion data in obstacle free situations with human guided trajectories and forces during obstacle avoidance. In addition, we use Differential Cooperative Game Theory (DCGT) based on human applied forces to integrate the two predictions, ensuring that the robot's behavior is consistent with human intentions. Compared with existing methods, DTRT incorporates human dynamics into long-term prediction, provides accurate intent understanding, achieves reasonable role allocation, and enhances the autonomy and operability of robots.

02. Physical human-machine collaboration faces dual challenges of intent prediction and role allocation

Physical human-machine collaboration (pHRC) is crucial in manufacturing, healthcare, and services. Effective strategies are needed to ensure that robots can seamlessly collaborate with humans, accurately estimate intentions, and dynamically adjust behavior to assist humans. Therefore, accurate human intent estimation and reasonable human-machine role allocation are key challenges in improving pHRC performance.

Accurate prediction of future trajectories based on human intent is crucial for the effectiveness of robot assistance and the safety of pHRC. In complex environments with potential hazards, such as when robots encounter unknown obstacles, the rapid changes in human intent pose significant challenges to intent estimation. The current methods mainly rely on short-term motion data, such as position and velocity, which limits the detection of human intention changes and will affect the accuracy and safety of predictions.

In addition, short-term data can reduce the effectiveness of intention estimation in long-term collaboration. Therefore, researching long-term prediction methods that integrate human applied forces into pHRC is crucial for achieving accurate intent estimation. In addition, human-machine role allocation involves a complex mechanism for assigning task control between humans and robots. This process will coordinate the human-machine relationship in real-time, reduce disagreements, and improve the robot's assistance level. The existing methods mainly rely on impedance/compliance control, and determine roles by modifying model parameters. Among various methods, role allocation based on game theory simulates the collaborative process among multiple participants, with the goal of minimizing the cost function to achieve optimal human-machine collaboration. However, ensuring that robot behavior is consistent with human intentions while maintaining autonomy and flexibility remains a major challenge.

03. DTRT exhibits significant advantages in prediction accuracy and collaborative performance

The experimental results show that DTRT has significant advantages in prediction accuracy, exhibiting excellent performance in multiple physical human-machine collaboration indicators, effectively reducing human-machine divergence and improving the robot's assistance level, with enormous potential for application. The core advantages of DTRT lie in the following three points:

DTRT combines human intention estimation with human-machine role allocation to detect intention changes and reduce human-machine divergence, effectively improving human-machine collaboration performance in complex and dangerous environments.

2. The hierarchical structure based human intent estimation in DTRT simultaneously processes motion and force data in human-machine collaboration, improving the prediction accuracy of human intent and providing accurate understanding of intent.

3. The human-machine role allocation based on differential cooperative game theory realizes adaptive leader switching based on the force exerted by humans, ensuring that the robot's behavior is consistent with human intentions and reducing divergence while maintaining robot autonomy.

In terms of physical human-machine collaboration performance, the following indicators were used:

In the experiment, we replaced and randomly located obstacles. The results indicate that DTRT outperforms existing methods in physical human-machine collaboration scenarios. The average human-machine collaboration angle of DTRT is 76.4 °, the average robot assistance level index is 1.5, and the human-machine system is in a collaborative state 61.8% of the time. In addition, the allocation of human-machine roles effectively balances the autonomy of robots and human guidance, resulting in only 3.5 J of human mechanical work. Overall, DTRT accurately estimates changes in human intent, reasonably allocates the roles of leaders and followers, thereby reducing human-machine divergence, improving the level of robot assistance, and promoting safe and efficient physical human-machine collaboration.

Quantitative comparison between DTRT and existing methods in terms of indicators

The proposal of DTRT is not only an algorithm breakthrough, but also an attempt to reconstruct the "human-machine relationship": it provides a technical path for the development of humanoid robots that combines universality and engineering value. In the future, the research ideas and core mechanisms of DTRT are expected to continue to expand and deepen in multiple practical application scenarios focused on by Lingbao CASBOT, such as industrial manufacturing, complex operations, and service collaboration.