Robotics / source study / edition in preparation
Seeing an object is only the beginning.
How detection outputs become a grasping decision. Read the mechanism on its original drawing.
Download the offline source study PDF ↗
US20240198530A1 · FIG. 2 · original PDF page 3. Rotated upright; geometry and labels preserved. Blue leaders identify parts and are not new process connections. Open the image to inspect it at full size.1208 / 210 · Detection modulesFind objects and possible grasps.
The example object detector locates objects of interest; the grasp detector supplies candidate grasp locations. The description also allows both inputs to come from one sensor.
Description [0027] · original page 8 ↗ 2212 · High-level sensor fusionRelate the outputs.
The fusion module combines detection outputs and computes attributes for the alternatives, including relationships between candidate grasps and detected objects.
Description [0027] · original page 8 ↗ 3220 / 218 · Objectives and constraintsSpecify what matters.
The ranking can reflect which objects to pick and constraints imposed by the environment. In the described embodiments these can be predefined or supplied by a user.
Description [0027] · original page 8 ↗ 4222 · Multi-criteria decision makingRank the alternatives.
Each criterion in the described decision matrix has a weight. A candidate is evaluated across those criteria before a grasp is selected. The patent gives several possible decision methods.
Description [0045] · original page 10 ↗ 5224 · ActionTurn the selection into instructions.
The selected alternative defines the output action. Executable code can then be sent to the robot controller to carry out the selective grasp.
Description [0047] · original page 11 ↗ The central reading
Detection supplies possibilities. Criteria guide the choice.
The disclosed system separates detection from the choice of action. It combines what the detectors report, attaches attributes to candidate grasps, then ranks the alternatives using the application's objectives and constraints. That distinction makes the decision criteria visible rather than treating a predicted grasp score as the whole decision.
What changes after an attempt?
Paragraph [0048] describes an optional way to adjust criterion weights using simulation or real-world results, within ranges initially defined by an expert. Improved settings can become the starting point for the next optimization step. This is a specified tuning procedure, not evidence of unrestricted learning or guaranteed safe performance.
Inspect paragraph [0048] on original page 11 ↗