System Development Review

Fall Semester System Review

Overview

This fall, our team took the autonomous seedling-planting system demonstrated at the Spring Validation Demonstration (SVD) and grew it into a fully integrated, multi-seedling planting pipeline. The Warthog base now drives to several seedling locations, avoids and replans around obstacles, digs with force feedback, plants with the xArm, and tamps the soil. An operator plans and starts each mission from a map-based user interface.

At the start of the semester we scoped down to focus on what matters most for the Fall Validation Demonstration (FVD):

  • Removed: pot detection and Cartesian path planning for the arm
  • Focused on: tamping, digging feedback control, obstacle avoidance, multi-seedling navigation, and a user interface monitor

Fall Semester Progress at a Glance

SubsystemStart of Semester (post-SVD)End of Semester (PR 11)
NavigationField-tested for a single seedlingMulti-seedling navigation, smooth turns, arrives within 1 m of each target
Obstacle AvoidanceGround segmentation untestedStops for obstacles and replans around static obstacles (Dijkstra on occupancy grid)
Seedling ManipulationDeterministic motion planning for one seedlingPick-and-place for 5+ seedling positions, integrated with navigation at full speed
Digging Feedback ControlForce sensors purchasedCompleted: threshold-based stop, retract, and recovery integrated into the planting FSM
TampingConcept sketchesFourth design iteration with a linear-motion linkage, field-tested
User Interface MonitorList of useful metricsInteractive Foxglove UI: select locations on a satellite map, send the plan, start the mission

Subsystem Details

Navigation

The semester started with navigation field-tested for a single seedling using the SVD software stack. Multi-seedling testing was briefly delayed by an e-stop battery issue.

  • Field-tested one-seedling and multi-seedling navigation with timed stops
  • Fixed heading oscillations (between −1.2 and 1.2 rad) and wobbling. The Warthog now makes much smoother turns and arrives within 1 m of every seedling location tested (3/3)
  • Field-tested navigation and planting together at multiple seedling locations
  • Fixed the robot spinning in place when the initial heading error was larger than 10°. The controller now updates the heading change when an IMU message is lost, re-checks the turn direction during point turns, and stops within 20° of the target heading
  • Integrated static obstacle replanning into the navigation stack

Obstacle Avoidance

We began by testing LiDAR ground segmentation in Schenley Park, then tested obstacle detection on Flagstaff Hill, where people standing around the robot were detected and marked on the costmap.

  • Verified that the robot stops in front of obstacles
  • Implemented path replanning around static obstacles using a Dijkstra search on the occupancy grid
  • Completed static obstacle replanning and integrated it with navigation

Seedling Manipulation

We mounted the end-effector on the xArm and added a dedicated battery pack so the arm runs on its own power.

  • Recorded waypoints for 5 seedling positions on the Warthog. All 5 seedlings were picked and dropped successfully with deterministic planning
  • Tested manipulation and navigation integration end to end
  • Found that the arm moved very slowly during integrated runs. Disabling the LiDAR nodes during planting helped reduce load but did not fix the problem on its own
  • Traced the slow motion to dropped DDS messages and fixed it (see System Integration), then added new pick-and-place positions

Digging Feedback Control

Digging feedback went from purchased parts to a completed subsystem this semester.

  • Bought force sensors (strip membrane FSR) and a shunt resistor, wired them up, and got initial analog readings
  • Ran bench tests for calibration and standing-load durability
  • Added a decoding feedback PDO to the LINAK actuator’s CAN node as a backup feedback source
  • Ran digging threshold trials in Schenley Park. In the failure recovery test, the auger stopped as soon as the threshold was reached and retracted to its initial position. Also tested the new LINAK position control
  • Updated the planting FSM so it reports COMPLETE, FAULT, or DIGGING_OBSTACLE. On a fault or obstacle, the actuators go to their home position and wait for the next step
  • Adjusted the code to work with the tamper’s motion. Tested indoors and outdoors and marked the subsystem complete

Tamping

We built the tamper in four iterations:

StageDesignResult
ConceptsThree options: (1) passive legs that pinch and compress soil when pressed down, (2) vertical motion that closes the arms, (3) a longer lever arm for more mechanical advantageBuilt small-scale prototypes
V1First full prototypeGathered some soil, but compaction was poor and the left and right sides were not synchronized
V22-way tamperTamped twice along each direction; next step was a 4-way design
V34-way tamperTested in Schenley Park. The soil there is much harder, the chute was not mounted firmly enough, and the tamper hung too low for ground clearance
V4Feet linkage changed from rotary to linear motionWorks, but results depend heavily on soil condition and the depth of the drilled hole

User Interface Monitor

We started by identifying the features and metrics worth monitoring and organizing all ROS 2 topics in a spreadsheet.

Made the UI interactive. The user can select planting locations on a satellite map, send the plan to the robot, and press “Start planting” to begin the mission

Built the first UI layout in Foxglove: robot location on the map, costmaps and obstacles, mission progress, and battery level

Added mission visualization on the map panel: the robot’s live position (yellow), its trail (blue), and the global path (orange), with color-coded seedlings for the current target (orange), planted (gray), and remaining (green)


System Integration

Most of the second half of the semester went into running all subsystems together. These were the main problems and how we handled them:

IssueRoot CauseResolution
Jetson overloadRunning all subsystems at once and recording every ROS 2 topicLimited publishing frequencies and sensor load, recorded an explicit topic list, and explored the MAXN (25 W) power mode. Resolved (PR 11)
Slow arm motionMotion planning was not the bottleneck. DDS dropped ROS 2 action messages between MoveIt and the arm controller and resent them only after 3 s, leaving the arm idle for minutesLarger UDP socket buffers on the Jetson and a shorter Fast DDS resend interval. Resolved (PR 11)
Behavior FSM bugThe plan manager removed two seedlings from the plan when only one had been planted, so navigation thought the mission was doneFixed during PR 10–11 integration
Robot spinning at startInitial heading error larger than 10° caused repeated point turnsHeading handling for lost IMU messages and repeated turn-direction checks. Resolved (PR 10)

Risk Management

RiskMitigation
Broken or flood-damaged parts (including a battery charger) slowing testingBought replacements
Noisy digging feedbackDeveloped two feedback methods: the FSR force sensor and LINAK CAN feedback
Breaking the auger during digging testsBought spare augers and adapters
Pittsburgh weather limiting outdoor testingIndoor unit tests and debugging with rosbags
Less sponsor support after a key contact leftSet up midterm and final meetings with Abhi, plus a shared Slack channel
Cable managementReorganized cabling, so the risk of damaging equipment is now low
Rain on demo dayFilmed full dry runs as a backup
Soil condition affecting tamping qualityMore testing to fine-tune the tamper