Skip to content

Tasks#

A task is one definition of a job for a vehicle. Each simulator tier uses the same definition. Each learner and each classical controller also uses it. Thus the scores of different methods are comparable.

Parts of a Task#

Part Content
Map products The geometry of the task. A script builds the products one time for each map version.
Mission The job, its completion test and its failure tests.
Plan The decisions that a planner makes before the vehicle moves.
Phases The sequence of steps in a mission. The task controls the sequence. A learned policy does not.
Actions The commands that a driver sends to the vehicle in each control step.
Observations The inputs that a driver gets in each control step.
Rewards The reward terms, their weights and the reasons for the weights.
Terminations The events that end an episode.
Metrics The scores of the evaluation.
Baselines The reference driver, the planners and their measured scores.

Location of the Code#

  • The code of a task is in Learning/acres_learn/tasks/<task>/.
  • The constants of the field scouting task are in Learning/acres_learn/configs/scouting_v1.json.
  • ACRES Core and the game read the same constants and the same map products.
  • The learners, the training and the deployment are on the Learning pages.

List of Tasks#

Task Vehicle Status Pages
Field scouting Polaris Ranger In use Field Scouting Task, Scouting Map Products, Scouting Baselines
Harvest of three corn fields Harvest machine Proposal Harvest Task

Field Scouting Pages#

  • Field Scouting Task is the specification. It gives the mission, the plan, the phases, the actions, the observations, the rewards, the terminations and the metrics. It has two formulations of the driving problem.
  • Scouting Map Products describes the geometry: ground classes, loops, checkpoints, entries and the transit graph.
  • Scouting Baselines describes the reference driver, the planners and their measured scores.

Two Formulations of the Scouting Task#

The mission, the map products, the plan and the metrics are the same in the two formulations.

Item Baseline Formulation Framework Formulation
Driver One flat PPO policy outputs the full command for the whole mission The reference driver gives the base command. A learned policy adds a bounded residual
Order of the fields The exact planner The exact planner. Lean proves that it is exact
Phase sequence The episode tracker The mission automaton. Lean proves its properties
Safety layer None. The reward has penalties A shield replaces each command that breaks a rule. Lean proves its rules
Reward v1 or bounded residual: the bounded reward, a shield term of 0.5 after the bound, a shaping term, failures at −1200 and a collision at −1500
Episode A stage of a curriculum One segment of a leg
State Exists. It is the negative baseline: 0 of 64 whole missions Specified, not built. The Lean models and proofs exist