Skip to content

Limitations#

This page lists the known limits of ACRES. Each model page has a section with the limits of that model.

Calibration Status#

Item Status
Polaris dynamics and drive-by-wire Fitted to the logs of the real vehicle.
Polaris LiDAR Calibrated against recorded scans of the RoboSense Helios.
Polaris camera The mount comes from a LiDAR and camera solution. The lab did not measure the mount on the vehicle.
Weather and soil water The Purdue Mesonet record of ACRE sets the conditions of a given day.
Maxxum and implements Not measured at ACRE. The values come from publications and datasheets.
The map Surveyed. The ACRE Scene gives the error budget.

The tyre, soil-tool, engine and fuel values of the Maxxum follow published data. No test compared the draft, the slip or the fuel of the simulated Maxxum with a real Maxxum on ACRE soil. Use the Maxxum results as engineering estimates.

Platforms#

  • The repository packages and tests the game on Linux with Vulkan and an NVIDIA GPU.
  • Windows: the commit history records one build, one package and one drive test. No sensor test on Windows exists.
  • macOS: the project has a configuration. No build exists.
  • The GPU LiDAR needs hardware ray tracing. Other machines use the CPU ray caster.

Platforms and GPU Tiers gives the requirements.

The Simulators#

Limit Effect
ACRES Core has the Polaris only. The Maxxum and the implements run only in the game.
ACRES Core has no camera. Use the render-later method to make images from a Core episode.
ACRES Core and the game use different rigid-body integrators. The trajectories agree within a measured spread. They are not identical.
The game is not fully deterministic outside lockstep. Frame timing changes the time at which a command applies.
One frame advances the physics by 0.1 s at most. A frame that is longer than 0.1 s makes the simulation slower than real time.
The farm traffic is kinematic. The traffic does not sink, slip or use fuel.
The optimizer gives steady-state estimates with the legacy soil model. The simulation can differ from a plan. In a test of two chisel-plow traction plans, the tractor stopped in the field.

Implements#

Limit Effect
Position mode holds the hitch angle relative to the tractor, not the depth below the ground. Pitch and sinkage change the ground depth. A repeat test held the angle at 8.457° while the depth reached 0.286 m and the engine fell to 603 rpm. Draft mode held 7.92 km/h in the same test.
The hitch in float mode can reach the lower stop of the links. The disc harrow and the seed drill then operate at a smaller depth than the setting.

Attach and Operate Implements gives the settings that operate correctly.

Sensors#

Limit Effect
The camera sensor model applies to the 8-bit image after tone mapping. The noise is not physically exact in linear light. No camera ISP exists.
The rolling shutter is a warp of one render for each row. Fast motion inside one frame is approximate.
A GPU with little memory uses the reduced sensor profile. The images differ from the calibrated profile. The dataset contains the label of the profile.
The local bags with LiDAR and camera of the real vehicle are short. The camera comparison uses few real frames.

Camera, LiDAR and INS and GNSS give the limits of each model.

ROS 2#

  • The workspace uses ROS 2 Humble from RoboStack in a conda environment.
  • The DDS loopback fence keeps the ROS 2 graph on one machine.
  • The vehicle package purdue_ranger of the lab is proprietary. The repository does not contain it.
  • The Polaris mesh for RViz is a reconstruction. It is not a measured model.

Build the ROS 2 Workspace shows what a user without the vehicle package can build.

The Real Vehicle#

WARNING

ACRES does not make the real Polaris safe. A policy that operates correctly in the simulator can fail on the vehicle. Only trained persons of the lab operate the vehicle, with the safety procedures of the lab.

  • The repository reads the logs of the real vehicle. It does not command the vehicle.
  • The field-day kit and the deployment program run on the vehicle only when the lab starts them.

Third-Party Assets#

Some models, textures and data have their own licences. The tractor body model permits no commercial use. README.md in the repository root lists each asset with its source and licence.