Truck Backer-Upper lab

Driving Field

x position
30.0 metres
y position
-8.0 metres
trailer angle
22.9 degrees
hitch angle
0.0 degrees
steering
0.0 degrees

Drag the trailer to move the truck · drag the cab to rotate it

  • Loading dock
  • Cab
  • Trailer
  • Front-wheel tire marks
  • Training episode path
  • Policy preview (dotted)

Status: Untrained. Enter an epoch count and press Train.

Neural Network Controller

Drive telemetry

Drive summary, No drive recorded yet. Run a drive to trace its hitch, steering and heading error.

Training Progress

No recorded training samples

Experiment controls

Number of training epochs (1 epoch = 4 episodes + 1 weight update)

Curriculum lesson that Random truck position samples from: the same start boxes training widens through, from short aligned approaches near the dock to the full-difficulty box: x 20 to 40 metres, any heading, hitch within 0.4 radians

train to capture checkpoints

Untrained. Enter an epoch count and press Train.

Saved experiments & scenarios

Save the current controller, training state, seeds and named poses in a JSON file. Loading a file replaces this session.


Field extents: x 0–50 m, y −25–25 m. Trailer heading and hitch angle are independent; terminal starts are allowed.

Named starting poses

Scenarios save the truck's current pose. Apply coordinate edits first. Reusing a name updates that scenario; save the experiment file to keep named poses between launches.


Policy evaluation

Map angles

These angles affect the position map only. Fixed benchmark starts stay the same.

Compare frozen controllers on the same starting poses.

Evaluation uses frozen controller copies and can run while training continues.

Compare the current controller and saved checkpoints on the same starts. Select a map cell after evaluation to inspect and replay that drive.

Backprop Through Time

Records one rollout and its backward sweep so you can step through how the terminal error flows back to the start.