Robotic Grasping is the subfield of robotics concerned with planning and executing stable, task-appropriate physical contact between a robotic end-effector and a target object. It integrates perception, grasp quality estimation, motion planning, and control to enable reliable object acquisition across varied geometries, materials, and environmental conditions.
Content
- Robotic grasping research began in the 1970s with analytical grasp mechanics studies by Salisbury and Craig, who formalised the conditions for force-closure (a grasp that can resist arbitrary external wrenches). Early industrial robots used fixed fixtures to avoid the grasping problem entirely — parts were pre-positioned to known locations. The 1980s and 1990s saw development of parallel jaw grippers with force sensing and initial work on form-closure and friction cone analysis. The introduction of RGB-D cameras (Kinect, 2010) and convolutional neural networks catalysed a shift from analytical to data-driven grasping, with GQ-CNN (Grasp Quality-CNN, Mahler et al., 2017) establishing a benchmark approach for planar grasp prediction from depth images.
- Modern robotic grasping systems follow a perception-planning-control pipeline. The perception module acquires a 3-D point cloud from depth cameras or stereo vision, segments the target object from clutter, and estimates its pose. The grasp planning module evaluates candidate grasp configurations — defined by contact points, approach vector, and gripper aperture — against quality metrics including grasp wrench space volume, antipodal score, or learned neural quality estimates. Motion planning finds a collision-free arm trajectory to the pre-grasp configuration. The controller executes the approach and closes the gripper with monitored force to confirm contact. Post-grasp verification uses force-torque sensors or tactile arrays to detect slip and trigger re-grasp if needed.
- Reliable grasping of novel objects in unstructured environments is economically transformative. Bin-picking — extracting randomly oriented parts from bins — is the canonical industrial application, with companies including Mujin, Covariant, and Righthand Robotics deploying learning-based grasping systems in fulfilment warehouses. Agricultural harvesting robots must grasp fragile, variably shaped produce without bruising. Surgical robotics requires micron-level precision grasping under endoscopic vision. Household service robots face the most demanding scenario: grasping arbitrary consumer objects from cluttered countertops — the “open world” problem that remains a significant research frontier.
- The 2024–2025 period has seen foundation models profoundly reshape robotic grasping. RT-2 (Google DeepMind, 2023) demonstrated that vision-language-action models pre-trained on internet-scale data could generalise grasping and manipulation to novel instructions and objects zero-shot. Diffusion-based action policies (Diffusion Policy, Chi et al., 2023) achieve state-of-the-art dexterity on multi-step manipulation tasks. Physical intelligence companies (Physical Intelligence / π, 2024) are building generalised robotic foundation models trained on diverse manipulation demonstrations. Dexterous hand hardware — including 16-DOF multi-fingered hands with embedded tactile sensor arrays — is becoming commercially available, enabling in-hand manipulation research that was previously confined to academic labs with bespoke hardware.