Dexterous manipulation is the ability of a robot to deliberately move, reorient or reposition an object through coordinated contacts, often while the object remains in the hand. It goes beyond simply picking something up: the robot may rotate a key, shift a tool between fingers, slide an object to a better contact point or regrasp it without putting it down.
This is difficult because every contact changes the physics of the task. Fingers can stick or slide, an object can pivot unexpectedly and small errors in force or pose can make a stable configuration fail.
Dexterous manipulation vs grasping
Robotic grasping focuses primarily on acquiring and stabilizing an object: choosing contact points, approaching, closing the gripper and maintaining a secure hold. Dexterous manipulation begins from a stable or partly stable grasp and intentionally changes the object’s pose or contact configuration. The two overlap, but their search and engineering questions are different.
In-hand manipulation
In-hand manipulation is a central form of dexterous manipulation. Instead of moving the whole robot arm to reposition an object, the fingers change their contacts so the object rotates or translates relative to the palm. Examples include turning a pen, orienting a small component for insertion or rotating a tool before use.
Rolling, sliding and regrasping
Contacts can change in several ways. A fingertip can roll over a surface while maintaining contact, slide intentionally, or release and establish a new contact in a regrasp. Successful manipulation often combines these modes. Planning them requires models of geometry, friction and reachable finger configurations.
Why dexterous manipulation is hard
A multi-finger hand has many joints, creating a high-dimensional control problem. The same object can be held in many valid ways, and each contact introduces constraints. Real surfaces also have uncertain friction, compliance and small geometric errors. A controller that works in simulation may fail when the physical object is slightly different.
Role of tactile intelligence
Tactile intelligence can provide the feedback needed to estimate contact changes that cameras cannot see. Slip, local pressure and force direction help the controller decide whether to tighten a grasp, move a finger or abandon a risky trajectory. Touch is particularly useful when fingers occlude the object.
Vision, proprioception and object state
Dexterous systems usually combine multiple sensing modes. Vision estimates object pose before and during manipulation. Joint encoders report finger configuration. Force and tactile sensors reveal contact. A state estimator combines these observations into a representation the controller can use.
Learning dexterous policies
Reinforcement learning, imitation learning and diffusion-based policies are increasingly used for high-dimensional manipulation. Simulation can generate large numbers of trials, while demonstrations provide realistic behavior. The challenge is transferring policies to hardware without relying on unrealistic contact physics.
Dexterous manipulation and VLA models
VLA models can help connect a natural-language goal to action sequences, but high-level action prediction does not remove the need for fast contact control. A practical architecture may use a large model for task planning while a lower-level policy handles tactile corrections and joint control.
Robot hands and end effectors
Dexterous manipulation can use anthropomorphic multi-finger hands, specialized grippers, soft hands or hybrid end effectors. More joints increase the possible motions but also increase sensing, actuation and control complexity. See dexterous robot hands for the hardware side of this problem.
How performance is evaluated
Useful metrics include task success, final object-pose error, time to completion, dropped-object rate, recovery from disturbances and generalization to new objects. Researchers also evaluate whether policies work across different friction conditions and whether contact forces remain within safe limits.
Why progress matters
General-purpose robots will need more than reliable pick-and-place. Home, warehouse and industrial tasks often require turning, inserting, folding, fastening, opening or using tools. Dexterous manipulation is therefore one of the major gaps between impressive robot demonstrations and flexible physical work.
