Skip to content

Harvest Task (Proposal)#

This page is a proposal for a second benchmark task: harvest three corn fields with the minimum fuel while rain occurs. The task does not exist in the code. All values on this page are proposal values.

Note

The proposal needs a combine harvester. The game does not contain one. The source model of a combine is in Acres/Raw/farm-equipment/combine_harvester. Roadmap lists the work. Tasks lists the tasks that exist.

Scenario#

Item Proposal
Start The machine is in lot A at the Beck Agricultural Center. The engine is at idle.
Fields F53, F54 and F49: one block of about 1.8 ha of corn, 250 m to 310 m from the lot.
Grain About 19.5 t at 1.1 kg/m².
Bins The grain bins in the west bin yard, east of F49. The unload pit is the only permitted place to unload.
Soil The SSURGO soil of the fields.
End The harvest of the three fields is complete, the tank is empty, and the machine is in lot A again. A time limit also ends the task.

Machine#

Parameter Proposal Reason
Corn head 6 rows of 0.76 m, 4.57 m wide The usual head of a mid-size combine.
Grain tank 8 t The job then needs two or three unloads. The time of each unload is a decision.
Unload rate 100 kg/s with the machine stopped at the pit A full tank takes about 80 s.
Harvest speed 3 km/h to 8 km/h with the header down
Road speed 25 km/h maximum with the header up

Weather#

Setting Proposal
Date and start time 20 September, 10:00.
Rain One event of 1 h with 5 mm to 25 mm. It starts when 30 % to 60 % of the job is complete.
Clock The farm clock can run faster than the machine, so the rain and the drying fit into one run.
Start wetness 0.3 to 0.8 of the range from the wilting point to the field capacity.

Rain changes the job in three ways. Wet soil gives more slip, more sinkage, more fuel for each hectare and deeper ruts. Wet grain has a drying cost. The visibility decreases. The agent can continue, drive slower, wait, or change the sequence of the fields.

Constraints#

The agent gets no reward for grain alone. These rules define a correct harvest. The task measures each rule at each step.

Number Constraint Measure Limit
C1 Follow the rows. The distance of the header centre from the planned pass, and the heading error. 0.25 m, 5°
C2 Complete coverage. The crop that stays in the fields. 2 % of the area
C3 Small overlap. The area that the header cut two times. 5 %
C4 Do not drive on crop. The crop that the wheels or the body crushed before the cut. 0.5 % of the area
C5 Header discipline. The header is down only in a field. No violation
C6 Speed limits. Harvest 3 km/h to 8 km/h, roads 25 km/h, yard 10 km/h. No violation longer than 2 s
C7 Tank discipline. Grain that the machine lost with a full tank or unloaded outside the pit. 0 kg
C8 Permitted ground. The machine stays on fields, farm roads, gravel and the yard. No violation
C9 Soil care. The area of ruts deeper than 5 cm on wet soil. Reported in m²
C10 Safety. Contact with a person or a vehicle. No contact. A contact ends the episode.

Objective and Reward#

The objective is to complete the job with the minimum fuel, without a violation of the hard constraints. The reward is the economic result of the job in dollars, so each weight has a unit and a source.

\[ r_t = p_g\,\Delta G_t - p_f\,\Delta F_t - p_d\,\Delta G_t\,\max(0,\, m_t - 15.5) - p_c\,\Delta A^{\text{crush}}_t - p_r\,\Delta A^{\text{rut}}_t - p_h\,\Delta t \]
Symbol Quantity Unit Proposal Value
\( \Delta G_t \) Grain that entered the tank in the step kg
\( p_g \) Grain price $/kg 0.18
\( \Delta F_t \) Fuel that the engine used in the step L
\( p_f \) Fuel price $/L 1.00
\( m_t \) Grain moisture, wet basis %
\( p_d \) Drying cost for each moisture point above 15.5 % $/kg 0.0016
\( \Delta A^{\text{crush}}_t \) Crop that the machine crushed before the cut m²
\( p_c \) Value of the lost grain $/m² 0.20
\( \Delta A^{\text{rut}}_t \) New rut area under the conditions of C9 m²
\( p_r \) Cost of compaction $/m² 0.05
\( p_h \) Cost of machine and operator time $/s 0.04

The constraints C1, C5 to C8 and C10 are separate cost signals with budgets. They are not reward terms. A large fixed penalty in the reward can teach an agent to stop and do nothing.

The benchmark score compares methods. Hard rules multiply. Quality terms give a weighted mean.

\[ \text{Score} = \text{coverage} \times \prod_{k \in \{C5,\, C7,\, C8,\, C10\}} g_k \times \frac{\sum_i w_i\, Q_i}{\sum_i w_i} \]
Symbol Quantity
\( g_k \) The gate of a hard rule: 1 without a violation, 0.5 to 0 with violations.
\( Q_i \) A quality term between 0 and 1: fuel for each hectare, row error, crushed crop, rut area, time.
\( w_i \) The weight of a quality term.

Difficulty Levels#

Level Content
L0 One field, dry, no tank limit.
L1 Three fields, start in lot A, the 8 t tank, dry.
L2 L1 with one hour of rain and a random start wetness.
L3 L2 with farm traffic, workers, and low sun or fog.
Held out A different block of corn fields with weather seeds that the training did not use.

Interfaces#

The proposal gives three interfaces to one task with one scorer.

Interface User Step Actions
Skills Language models and planners On events and each few simulated minutes Tool calls such as drive_to, harvest_pass, unload and wait. Controllers of the task execute them.
Waypoints Planners with their own path tracking 1 Hz to 2 Hz A path or a target pose, a speed, and the modes of the header and the auger.
Control Policies 10 Hz Steering, pedal, gear, header, auger and differential lock.

All interfaces give the same observations: the front camera, a job map image, the state vector, events and the used budget.

What Exists Now#

Need of the Proposal Status
A fuel model in the vehicle state and in the session log Exists. Refer to Energy and Fuel. The CAN bus has no fuel frame.
Lockstep with deterministic steps Exists. Refer to Lockstep Stepping.
Rain, soil water and their effect on traction Exists. Refer to Soil Water.
The harvest of crop by a cutter Exists for the 3 m cutter on the Maxxum. Refer to Crops and Ground Classes.
A combine harvester with a grain tank and an unload auger Does not exist.
The constraint monitors, the reward, the score and the report Do not exist.
A reference planner and a scripted driver for the harvest Do not exist.