Grasping from a single view
6-DoF grasp detection for parallel-jaw grippers in cluttered scenes, from one RGB image or one RGB-D view.
A robot picking objects from a cluttered table usually sees the scene from a single viewpoint. Much of each object is hidden: its back, its underside, the parts pressed against its neighbours. A parallel-jaw gripper still needs a complete 6-DoF pose that closes on the object without colliding with anything else. My PhD research at Swansea University studies grasp detection under exactly this condition.
Recovering hidden geometry from one RGB image
Most grasp detectors rely on depth input or on 3D models of the objects. I asked how far a single monocular RGB image can go. Triplane Grasping reconstructs an object with a hybrid triplane-Gaussian representation and then predicts 6-DoF grasps by treating the reconstructed points as candidate contacts (ICARM 2025). AAO-Grasp extends this to cluttered scenes: a semantically guided diffusion model builds a scene template from the image, and a coarse-to-fine structure resolves scale at the scene and object levels. Grasps are then generated over the reconstruction, which also allows grasp prediction from unseen viewpoints (RoboticCC 2025).
Modelling the ambiguity that remains
Even with depth, one view leaves grasp estimation ambiguous: sensor noise corrupts the measured geometry, object sizes vary widely, and a rotationally symmetric object admits several equally valid grasps. ACE-Grasp treats grasp estimation as uncertainty-aware distributional modelling instead of deterministic regression. It rectifies unreliable 3D geometry with 2D semantic consistency, samples approach directions from a learned distribution over a dense set of views, predicts the in-plane rotation as a distribution rather than a single value, and aggregates multi-scale features with a dynamic receptive field (ICPR 2026).
Tools and data
To run published detectors under one evaluation and to test components across them, I develop GraspPanda, an open-source toolbox for 6-DoF grasp detection in clutter. I also coordinated the construction of RASD, a dataset for segmenting robotic arms in complex scenes (RoboticCC 2025).