Dynamic Control of Nonprehensile Manipulation Systems

Nonprehensile manipulation refers to robotic interaction methods that move objects without relying on secure grasps, instead exploiting dynamic contacts such as pushing, sliding, rolling and flicking. Recent advances have centred on dynamic control—strategies that modulate robot actuation in time to shape object trajectories through controlled contact forces and kinematics. These approaches extend beyond quasi-static pushing to leverage frictional interactions, transient impacts and inertial effects. Key research themes include the synthesis of manipulation primitives that combine rolling and sliding to reorient objects, the integration of high-bandwidth tactile and visual sensing for real-time feedback, and the deployment of learning-based methods to acquire control policies capable of coping with uncertain contact dynamics. Practical applications span industrial in-plant logistics involving throwing and catching tasks, robotic assembly lines performing precise sliding operations, and service robots handling delicate items in unstructured environments. By embedding dynamic control principles into both model-based frameworks, such as model predictive control, and data-driven paradigms, particularly deep reinforcement learning, the field is converging on robust, adaptable solutions for complex nonprehensile tasks. This synergy of theory and experiment is elevating robotic dexterity, enabling machines to manipulate a wide variety of objects at high speed and with minimal computational overhead.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Dynamic Control of Nonprehensile Manipulation Systems publication trend

The graph below shows the total number of articles in dynamic control of nonprehensile manipulation systems across all publications each year (not limited to Nature Index journals).

Technical terms

Nonprehensile manipulation: Object handling by means other than secure grasps, e.g. pushing, sliding or rolling physical contacts.

Dynamic control: Time-varying actuation strategies designed to exploit object inertia and contact dynamics for manipulation tasks.

Model predictive control (MPC): An optimisation-based method that uses a predictive model to compute control inputs by minimising a cost function over a future time horizon.

Deep reinforcement learning (DRL): A machine learning approach where control policies are learned through trial-and-error interactions with an environment, guided by reward signals.

References

1. Multi-camera tracking of mechanically thrown objects for automated in-plant logistics by cognitive robots in Industry 4.0. The Visual Computer (2024).
2. Dynamic Nonprehensile Manipulation of a Moving Object Using a Batting Primitive. Applied Sciences (2021).
3. Robot Anticipation Learning System for Ball Catching †. Robotics (2021).