Introduction
Assistive robotics, particularly those designed for individuals with motion impairments, play a crucial role in restoring independence for activities of daily living (ADLs) such as drinking. However, traditional robotic manipulators often move in ways that are mathematically optimal but feel "unnatural" or unpredictable to human users. In Human-Robot Interaction (HRI), predictability is essential for user comfort. This research presents a self-supervised learning framework that enables robot arms to mimic the fluid, curved trajectories of human motion, enhancing the user experience in assistive contexts.
Methodology
The proposed framework utilizes a self-supervised imitation learning approach to translate human motion into robotic trajectories. The process is divided into three primary stages: data capture, representation learning, and deployment.
Data Capture and Representation
Instead of relying on complex vision systems, the researchers used three wearable Inertial Measurement Unit (IMU) sensors placed on a subject's upper arm, forearm, and hand. This setup allows for "action-free" demonstrations, capturing the natural dynamics of the task without requiring the robot to be present during initial data collection.
Architecture: STS-GCN
The core of the system is a Space-Time Separable Graph Convolutional Network (STS-GCN). This architecture treats the human arm joints as nodes in a graph, allowing the model to learn spatial relationships between joints and temporal relationships across frames independently. The model was modified to be autoregressive, enabling it to recursively generate a 120-frame trajectory based on its previous outputs. Training was conducted using Mean Per Joint Position Error (MPJPE) to minimize the discrepancy between predicted and actual human paths.
Experimental Setup
The study involved five participants performing drinking tasks. Each participant provided 10 demonstrations across six different bottle positions arranged in a 2x3 grid. The resulting trajectories were down-sampled and segmented into three distinct phases: reaching, drinking, and returning. The generated wrist trajectories were then linearly scaled and mapped to the workspace of a UR5e robot arm.
Results and Discussion
The system successfully generated functional, human-like drinking trajectories for the UR5e robot. A key finding was the model's ability to capture "hysteresis"—the phenomenon where the forward path of a motion differs from the return path. While standard Inverse Kinematics (IK) typically produces linear, rigid paths, the GNN-based approach naturally reproduced the biological "arc" characteristic of human movement.
Qualitative comparisons demonstrate that the GNN-generated paths are significantly more curved and fluid than traditional joint-space trajectories. Furthermore, the model showed impressive data efficiency, learning effective representations from only 50 total demonstrations. This makes the approach highly practical for real-world applications where large-scale data collection is often unfeasible.
Conclusion and Future Work
This work demonstrates that representation learning via GNNs can effectively bridge the gap between human motion and robotic control, creating more predictable and comfortable assistive interactions. Current limitations include the assumption of static environments and a lack of real-time visual feedback. Future research will focus on integrating computer vision to handle dynamic obstacles and incorporating force-sensing to manage physical interactions between the robot and the user during complex ADLs.
Summary updated