Project overview
The work introduces CSA-AInceptNet, a unified architecture for two tightly coupled manipulation tasks: identifying an object and estimating where it should be grasped. Channel sharpening attention and adaptive inception processing are used to strengthen shared event-based features while transfer learning improves generalisation across datasets.
Core contributions
- A single multi-task model for event-based object classification and grasping-position detection.
- Channel sharpening attention to emphasize discriminative event features.
- Adaptive inception processing to capture information at multiple spatial scales.
- Cross-dataset transfer learning from E-Grasp to Neuro-Grasp for robotic manipulation.
Reported evidence
On Neuro-Grasp, the transferred model reports a mean IoU of 0.4897 in addition to 98.58% classification accuracy. Comparative and ablation experiments evaluate the proposed architecture against ConvNeXt, DarkNet, DenseNet, VGG16, and other alternatives.
Datasets and evaluation
- E-Grasp: primary event-camera benchmark for object recognition and grasp-position detection.
- Neuro-Grasp: target dataset used to evaluate transfer learning and cross-dataset generalisation.
- Joint evaluation using classification accuracy and mean Intersection over Union for grasp localization.
Relevance to human-in-the-loop robotics
Joint recognition and grasp detection can reduce the number of separate perception pipelines required by a robot. This is especially useful in dynamic human–robot workspaces, where the robot must quickly understand both what an object is and how to interact with it.