Reactive Motion Generation and Control of ElasticJoint Robots for Reaching and Grasping

Kurzfassung

In the human motor system, the interplay between musculoskeletal elasticity and neural reflexes enables seamless adjustments during interaction tasks such as object reaching and grasping. This synergy unifies motion planning and control into a single, reactive process that ensures adaptability and fluidity. In contrast, robotic hand-arm systems typically rely on a decoupled approach, where arm and finger motions are planned separately within a "Sense-Plan-Act" paradigm. This leads to sequential, open-loop execution, which limits reactivity and results in less natural movement. The high dimensionality of multi-fingered hands further complicates real-time motion planning, while achieving effective compliant control in robots with elastic joints remains a significant challenge. Consequently, robotic grasping often appears rigid and brittle, struggling to cope with dynamic environments or produce adaptive, human-like behaviors. This thesis addresses these challenges by exploring the integration of reactive, compliant, and dynamic control into robotic reaching and grasping, inspired by the remarkable adaptability of the human motor systems. The work makes three primary contributions to advance the field of elastic-joint robotics.

To emulate the role of reflexive control in the human motor system, the thesis begins by developing specialized control laws for antagonistic actuators, such as bidirectional variable-stiffness designs and unidirectional tendon-coupled systems, commonly used in robotic arms with multi-fingered robotic hands. Building on the original Elastic Structure Preserving impedance (ESPi) framework, the extended control law preserves the intrinsic elasticity of these actuators while enabling user-defined impedance behavior. The closed-loop system ensures global asymptotic stability and output-strict passivity, providing a robust foundation when executing compliant and responsive reaching and grasping motions. Leveraging the compliant and stable foundation established by the extended ESPi control framework, the thesis introduces Finger Flow (FF), a real-time vector-field-based motion generation framework that utilizes real-time visual and proprioceptive feedback into continuous velocity commands. A computationally efficient geometric abstraction for hands with high degrees of freedom enables real-time finger-avoidant reaching motion. This empowers robots to grasp unpredictably moving cylindrical objects in free space. After a grasp is established, the thesis proposes a task-space motion and force tracking controller that preserves the robot's intrinsic elastic structure during constrained interactions. Preserving this structure maintains natural compliance and ensures safe, stable interaction. Unlike conventional approaches that enforce full decoupling of motion and force subspaces, the method avoids canceling out the intrinsic dynamic coupling between them. This prevents potential instability in compliant systems due to the modeling errors. Instead, the controller ensures passive closed-loop dynamics and global asymptotic stability, enabling simultaneous tracking of desired motion and force while respecting the robot's physical compliance.

Together, these contributions form a reactive "Sense-Act" paradigm that enables smooth, continuous transitions between reaching, grasping, and manipulation through reactive control. Experiments and simulations on the DLR robot neoDavid validate the approach, demonstrating the practical viability of preserving physical dynamics while delivering robust performance in dynamic, human-shared environments, enabling more natural, adaptive, and safe robotic manipulation.