Inside Loomo.launch [xml], users can change the most important parameters of the pipeline depending on the requirements of the test, on the server where is being launched, ...
See the tradeoff section at the end.
| Parameter | Description |
|---|---|
| ip_address_robot | IP Address Autonomous System (client) |
| ip_address_neuro (not used) | IP Address 2nd Client: NeuroProstethics Device (keypoint information) |
| v_max | Robot's maximum speed [m/s] |
| wheel_base | Robot's length between wheels [m] (0.57 for Loomo) |
| speed | Constant speed of the robot [m/s] |
| Parameter | Description |
|---|---|
| PERCEPTION_FUNCTION | Detection algorithm used: Stark |
| dt_perception | Perception minimum sampling time [s] (will just wait to finish if its above) |
| downscale | Downscale for robot images (check follow.cfg) |
| detector_size | DSize of detection model (small, medium, large) |
| tracking_confidence | treshold for tracking confidence 0.8 (id switches) -> 0.99 (not able to track) |
| keypoints_activated | Boolean to activate 3D Pose Estimation |
| save_keypoints_vid | Create a mp4 video to view 3D keypoints |
| keypoints_logging | Saving the coordinates of the 3D keypoints in a CSV file |
| visualization_percep | Boolean to visualize the image of loomo and bounding box |
| visualization_3D_activated | Boolean to visualize the 3D Pose in Real Time |
| verbose_percep | 0 (minimal) ->4 (full) level of verbose, above 2 to see runtime |
| Parameter | Description |
|---|---|
| ROBOT_STATE_FUNCTION | Robot State algorithm used |
| dt_robot_state | Robot State sampling time [s] |
| Parameter | Description |
|---|---|
| MAP_STATE_FUNCTION | Map State algorithm used |
| mapping_activated | Boolean to decide if we map the environment |
| map_state_activated | Estimate the state with map? |
| dt_map_state | Map State sampling time [s] |
| verbose_map | Boolean to indicate if we want to print messages |
| Parameter | Description |
|---|---|
| PREDICTION_FUNCTION | Prediction algorithm used |
| prediction_activated | Prediction algorithm required? |
| model_prediction_path | global path to the prediction model |
| time_horizon_prediction | Last predicted time [s] |
| past_observations | Number of past observations needed |
| dt_prediction | Prediction sampling time [s] |
| Parameter | Description |
|---|---|
| PATH_PLANNING_FUNCTION | Path Planning algorithm used |
| dt_path_planning | Path Planning sampling time [s] |
| time_horizon_path_planning | Last planned time for path calculation [s] |
| planner_type | Straight/Curvilinear path planning |
| robot_position | Right/Left relative to the patient |
| goal | Goal (detection collision avoidance) |
| workarea | Boundaries (detection collision avoidance) |
| Parameter | Description |
|---|---|
| CONTROL_FUNCTION | Control algorithm used |
| time_horizon_control | Last control MPC time [s] |
| n_states | Number of states for control |
| dt_control | Control sampling time [s] |
| Parameter | Description |
|---|---|
| VISUALIZATION_FUNCTION | Visualization algorithm used |
| dt_visualization | Visualization sampling time [s] |
When tuning the parameters of the program one must be conscious about the coupling effects of some parameters. The dt_control is one of the most important one and should be approximately the same value as the runtime printed on the Loomo app so that motion planning and communications can be synchronized. It is intuitive that adding more components to the perception node will slow it's inference therefore it's important to adapt the speed of the robot to the dt_control value.
Some combinations of parameters tht seemed to work for me:
- Simple tracking: speed up to 0.7 m/s, keypoints not activated, dt_control around 0.25s
- 3D keypoints: speed around 0.3 m/s, keypoints and visualizations activated, dt_control around 0.4s