Regrasping

Regrasping is not a random release; it is a deliberate, goal-oriented action sequence. The primary objectives are:

• Achieving Force Closure: Transitioning from an unstable, partial grasp to a configuration that can resist external wrenches (forces and torques).
• Functional Reorientation: Repositioning the object within the hand to a specific pose required for a downstream task, such as inserting a peg or using a tool.
• Overcoming Kinematic Limits: Moving the object to a new location when the robot's arm or wrist cannot directly achieve the desired end-effector pose due to joint range or obstacle constraints.

The defining action of regrasping is the temporary, controlled release of the object. This distinguishes it from in-hand manipulation, where the object remains continuously grasped. The release can be:

• Partial: One or more fingers open while others maintain contact, allowing the object to pivot or slide.
• Complete: All contacts are broken, and the object may be transferred to a secondary surface (like a table) or re-grasped mid-air.
• Supported: The object is often placed against a support surface (e.g., a tabletop) during the transition to provide stability and reduce planning complexity, a strategy known as extrinsic dexterity.

Successful regrasping requires tight integration across multiple planning and control layers:

• High-Level Task Planning: Decides when and why to regrasp within a larger activity graph.
• Grasp Sequence Planning: Computes the feasible intermediate and final grasp poses, often using grasp wrench space analysis or learned models like DexNet.
• Motion Planning: Generates collision-free arm and finger trajectories to execute the release and re-acquisition, potentially using Rapidly-Exploring Random Tree (RRT) or trajectory optimization.
• Low-Level Control: Employs impedance control or admittance control to manage contact forces during placement and pickup, and may use tactile servoing for precise alignment.

Regrasping in unstructured environments depends critically on perception to recover from uncertainty:

• 6D Pose Estimation: Continuously tracking the object's position and orientation after release, especially if it moves during the transition.
• Tactile Feedback: Using sensors like GelSight or force-torque sensors to confirm successful placement, detect slip during re-acquisition, and measure contact geometry.
• Visual Servoing: Employing camera feedback to guide the end-effector to the precise pre-grasp pose for the new grip.
• State Estimation: Fusing proprioceptive and exteroceptive data to maintain a belief over the object's state throughout the manipulation sequence.

Learning and planning regrasping policies is heavily reliant on simulation due to the high cost of physical trial-and-error. This introduces the sim-to-real gap.

• Physics Modeling: Accurate simulation of contact dynamics, friction, and object deformation is crucial but difficult.
• Domain Randomization: Varying simulation parameters (object mass, friction coefficients, sensor noise) is a key technique to train policies that transfer to physical robots.
• Benchmarking: Standardized object sets like the YCB Object Set are used to evaluate regrasping algorithms across different research groups.

Advancements in several adjacent technologies directly improve regrasping capability:

• Underactuated & Adaptive Hands: Robotic hands with fewer actuators than degrees of freedom can conform to objects, simplifying the re-acquisition process.
• Series Elastic Actuators (SEAs): Provide compliant force control for gentle placement and robust grasping.
• Contact-Implicit Trajectory Optimization: Planning frameworks that can automatically discover contact sequences, including regrasps, without manual specification.
• Model Predictive Control (MPC): Enables real-time adjustment of the regrasping motion based on the latest sensor feedback and predictions.

In automated distribution centers, robots use regrasping to handle a vast array of product shapes and sizes. A robot might initially grasp a box from the top, then regrasp it from the sides to stow it efficiently on a shelf or place it into a shipping container. This is essential for non-prehensile manipulation tasks like sliding or pivoting items into tight spaces. Systems like DexNet provide the underlying models to evaluate and select these alternative grasp poses from sensor data.

During complex assembly, a part must often be reoriented to align with another component. A robot arm might pick up a gear, regrasp it to flip it 180 degrees, and then insert it into a housing. This requires precise 6D pose estimation of both the part and the target, coupled with contact-implicit trajectory optimization to plan the release and reacquisition motions without collisions. Impedance control or admittance control is often used during the regrasp to manage contact forces safely.

A service robot preparing food must manipulate tools and ingredients with human-like adaptability. To use a spatula, it might grasp it by the handle, perform a task, then regrasp it closer to the blade for a scraping motion. This involves in-hand manipulation to adjust the grip initially, followed by a full regrasp. Tactile servoing with sensors like GelSight provides the feedback needed to detect slip during the transfer and ensure a stable new grasp.

In manufacturing kitting, robots retrieve parts from cluttered bins. An initial grasp might only be possible on an easily accessible but non-functional part of an object. The robot will then execute a regrasping sequence—often against a fixed surface like a table—to achieve a force closure grasp suitable for the next task, such as insertion. This process relies heavily on visual servoing and real-time trajectory optimization to execute the intermediate placement.

Surgical assistants may need to hand off instruments to a surgeon or reposition a tool within the operative field. A regrasping maneuver allows the robot to release a scalpel at a specific orientation for the human to take, or to adjust its hold on an endoscope for a better camera angle. These systems employ virtual fixtures to constrain motion to safe volumes and use proprioceptive sensing for extremely precise, jitter-free movements during the handoff.

Training regrasping policies is a major focus in robotics research. Using domain randomization in physics simulators, researchers train policy networks to perform regrasps under varied conditions (slippery objects, poor lighting). The policy learns to evaluate the grasp wrench space of its current hold and decide when a regrasp is necessary for stability. Bridging the sim-to-real gap for such contact-rich skills is an active challenge, often addressed with robust model predictive control (MPC) frameworks.

In-hand manipulation is the fine-grained control of an object within a robotic hand's grasp, using finger motions to reposition or reorient the object without releasing it. This is often contrasted with regrasping, which involves a release-and-reacquire cycle.

• Key Distinction: Maintains continuous contact and force closure.
• Techniques: Includes finger gaiting, rolling, and sliding.
• Purpose: Achieves precise object alignment for subsequent tasks like insertion or tool use.

The grasp wrench space is the set of all possible wrenches (combined forces and torques) that can be applied to an object by a robotic grasp through its contact points. It is a fundamental analytical tool for evaluating grasp stability.

• Calculation: Derived from contact point locations, friction cones, and applied contact forces.
• Use in Regrasping: A regrasp is often planned to transition an object from a grasp with a small, marginal wrench space (vulnerable to disturbances) to one with a large, robust wrench space.
• Force Closure: A grasp achieves force closure if its wrench space can resist any external wrench, meaning the origin of the wrench space is inside the convex hull of possible wrenches.

Tactile servoing is a closed-loop control method that uses real-time tactile sensor feedback to guide robotic manipulation. It is critical for executing regrasps in unstructured environments where visual occlusion occurs.

• Feedback Signal: Uses data from sensors like GelSight or force-torque sensors to measure contact geometry, pressure distribution, and slip.
• Regrasping Application: Guides finger placement during the reacquisition phase to ensure proper contact location and force. Enables correction for pose estimation errors.
• Contrast with Visual Servoing: Operates when the object is within the gripper and may be visually obscured.

Non-prehensile manipulation is a class of robotic manipulation where objects are moved without a firm enveloping grasp. Regrasping sequences frequently incorporate non-prehensile actions as intermediate steps.

• Common Actions: Pushing, pivoting, tumbling, or batting.
• Role in Regrasping: Used to reposition an object on a table (e.g., via a push) to make a desired grasp feasible. Expands the set of possible regrasping strategies beyond pick-and-place.
• Advantage: Often requires less precision than prehensile grasping and can handle objects too large or heavy to lift.

Task and Motion Planning is a hierarchical approach that decomposes high-level goals into sequences of feasible movements. It is the overarching framework that typically decides when and how to regrasp.

• Integration: Combines symbolic task planning ("pick up screwdriver", "use screwdriver") with continuous motion planning and collision checking.
• Regrasp Planning: The TAMP solver identifies that a current grasp (e.g., power grasp on screwdriver handle) is incompatible with a subtask (e.g., inserting tip into screw head) and inserts a regrasp action (e.g., to a precision grasp) into the plan.
• Solvers: Often use Rapidly-Exploring Random Trees (RRT) or optimization-based methods in the motion layer.

Slip detection is the process of identifying when an object begins to move relative to a robotic gripper. It is a critical failure mode sensor that can trigger an emergency regrasp.

• Sensors: Primarily uses tactile arrays or force-torque sensors to detect high-frequency vibrations or unexpected force changes.
• Reactive Regrasp: Upon detecting incipient slip, a controller can increase grip force or execute a fast, stabilizing regrasp to prevent total loss of the object.
• Preventive Role: Informs grasp quality assessment, indicating when a planned regrasp to a more stable configuration is necessary.