ROBOTCORE Collaborative

Our system captures depth images of the robot and its surroundings, requiring calibration of the relative positions between the robot and the camera. The robot's geometry is filtered out, and critical distances from control points on the robot to the surrounding depth camera detections are identified. Control points are defined using TF frames on the robot's geometry, typically on the axes connecting the joints. These points can be added by including link elements without geometry to the robot's URDF/XACRO model. The distribution of these points should ensure the entire robot is covered within the repulsive potential field's saturation region. The distance between consecutive points should generally be less than \(\left(\frac{-3}{\alpha} + 1\right) \frac{\rho}{2}\). More points increase precision but also computation time. Based on these distances, a Cartesian repulsive vector and an attractive configuration (PID) vector are computed. These vectors are combined in joint space to obtain a velocity command. The repulsion intensity (\(v_{\text{rep}}\)) is regulated by a curve defined by \(v_{\text{max}}\), \(\rho\), \(d_{\text{min}}\), and \(\alpha\). This velocity command is limited based on joint velocity limits from other control points and smoothed through an acceleration filter before being transmitted to the robot.

1. A stream of depth images is captured, showing the target robot and its immediate surroundings. The relative positions between the robot and the camera must be calibrated.
2. The robot's geometry is filtered out from these images.
3. The filtered image is analyzed to identify critical distances from control points defined on the robot's geometry to the detections from the surrounding depth camera. An example of these control points for a UR10e robot can be seen in the following image. These control points are defined using TF frames located on the robot's geometry, typically on the axes connecting the joints. These points can be simply defined by adding a series of `link` elements without associated geometry to the robot's URDF/XACRO model. The process of adding/modifying these control points to the configuration of the various involved nodes and the controller will be reflected in the documentation provided with the solution code. The distribution of these points along the robot's geometry should generally be sufficient to ensure that the entire volume of the robot is covered within the saturation region of the repulsive potential field function. This function appears in the next step, but as a rule of thumb, the distance between consecutive points should generally be less than \(\left(\frac{-3}{\alpha} + 1\right) \frac{\rho}{2}\). It is worth noting that the greater the number and density of these points, the more precise the coverage of the geometry at the cost of increased computation time.
4. Based on these distances for the end effector control point, a Cartesian repulsive vector is computed, and based on the current goal, an attractive configuration (PID) vector is also computed. These vectors are combined in joint space to obtain a velocity command. The intensity of the repulsion (\(v_{\text{rep}}\)) is regulated by a curve defined by the attached image: Where \(v_{\text{max}}\) is the maximum Cartesian repulsion velocity, \(\rho\) is the radius around a control point where obstacles are considered, \(d_{\text{min}}\) is the minimum distance from the control point to the obstacle, and \(\alpha\) is a shape factor that regulates the slope of the curve.
5. This velocity command is then limited based on velocity limits for each joint computed from the other control points.
6. The joint command limited to these new limits is smoothed through an acceleration filter.
7. This smoothed command is transmitted to the robot.