High-Speed Forward and Reverse Tractor-Trailer Driving with Learned Vehicle Dynamics

Aaron Chen*,1, David Shen*,1, Taekyung Kim*,2, Kaleb Ben Naveed2, and Dimitra Panagou2,3

University of Michigan  ·  *Equal contribution
1Electrical Engineering and Computer Science  ·  2Robotics  ·  3Aerospace Engineering

High-speed reverse driving  ·  Hidden vehicle dynamics  ·  Snow and uneven terrain

TL;DR

We learn tractor–trailer dynamics from recent motion to improve control when friction and terrain are unknown. In BeamNG.tech, we test reverse driving at target speeds of 30 and 50 km/h and forward driving at 80 km/h.

Learned MPPI completes 62 of 72 reverse trials, compared with 41 of 72 for analytic MPPI, with the improvement coming from fewer stalls. Both complete all 36 forward trials. The learned controller also transfers to slopes and banking absent from training, though reverse stalls remain a challenge on snowy slopes.

Control objective

Track the target speed, stay on the road, and avoid jackknifing.

v★

Track the target speed

Reach the commanded forward or reverse velocity without stalling.

α

Suppress hitch error

Keep the tractor–trailer articulation bounded to avoid jackknife.

H = 4

Use recent motion

Condition predictions on observed history without explicit friction or terrain inputs.

Why high-speed reverse driving?

A passive trailer is difficult to stabilize in reverse. Snow and uneven terrain make accurate motion prediction harder.

01

Coupled motion

The tractor controls a passive trailer through the hitch. A steering correction changes both vehicles’ motion.

02

Reverse instability

Small motion errors can grow into a jackknife. Higher speed leaves less time to correct them.

03

Unobserved conditions

Friction, slope, banking, and suspension affect the vehicle’s response but are not direct inputs to the learned model.

Much of tractor–trailer control research focuses on parking and low-speed navigation. We study reverse driving near handling limits, where errors in the dynamics model directly affect control.

A reverse trajectory on a snowy hill ends in jackknife under analytic-prior control.
Analytic prior. The trailer jackknifes in this example.
The learned controller keeps the tractor and trailer aligned during a reverse maneuver on the snowy hill.
Learned dynamics. The controller maintains alignment on the same hill.

What we evaluate

01

High-speed forward and reverse control

History-conditioned dynamics within MPPI for reliable tracking across four environments.

02

Prediction under uncertainty

Five-second rollouts from learned dynamics and analytic priors with and without recorded contact-friction information.

03

History length

Selected predictors with one, four, and eight steps of history on a separate Snow + Slope segment.

MPPI with learned dynamics

A neural network predicts vehicle motion from a short state-action history. MPPI uses these predictions to choose steering and throttle/brake commands.

MPPI initializes each rollout with four measured observations and three previously applied commands. The learned model predicts the next state given the rollout history and candidate command. MPPI uses rollout costs to update its plan and applies the first command in BeamNG.tech.
Controller architecture. Each rollout starts with four measured observations and three previously applied commands. The predictor takes this history and a candidate command, then advances the rollout. The applied command and next observation update the measured history for replanning.

Network

The predictor is an MLP with four hidden layers of 128 units each.

Inputs and outputs

The model uses hitch angle, velocities, yaw rates, actuator responses, and commands to predict accelerations and actuator rates. Kinematics update position and heading.

Training data

Training runs used flat asphalt at five uniform friction levels. Forward runs included weaving and command noise. Additional reverse runs covered larger hitch angles.

For comparison, we use a planar tractor–trailer model with Fiala tire forces.

Four test environments

We add snow and terrain changes separately to a Barcelona-Catalunya track layout in BeamNG.tech. Slope and banking are absent from training. The same snow pattern and terrain are reused in the combined setting.

Learned reverse-driving trajectory on a flat asphalt track.
(a) Asphalt. Flat pavement with uniform high friction.
Learned reverse-driving trajectory through snow patches on flat pavement.
(b) Snow. Flat pavement with low-friction patches.
Learned reverse-driving trajectory on sloped asphalt.
(c) Slope. Asphalt with elevation changes and banking.
Learned reverse-driving trajectory on snowy, sloped terrain.
(d) Snow + slope. Snow patches combined with uneven terrain.
Snow-plus-slope track map with snow coverage, elevation contours, track grade, and three spawn locations.
Track map. Snow coverage, elevation contours, road grade, and three starting locations.

Closed-loop performance

Each speed and environment has nine trials: three seeds at three starting locations. Completion means reaching 500 control steps (25 seconds), not finishing a lap.

More reverse completions

Learned MPPI improves completion in seven of eight reverse configurations, with 21 fewer stalls across all reverse trials.

Better speed and hitch tracking

Both MPPI variants complete all 36 forward trials. Learned MPPI brings mean speed closer to target and lowers RMS hitch angle at every forward and reverse setpoint compared with analytic MPPI.

Unseen terrain

Learned MPPI completes 16/18 reverse trials on Slope and 10/18 on Snow + Slope. Snowy slopes remain the hardest setting.

At the 50 km/h reverse target on Asphalt, learned MPPI averages 47.2 km/h and completes all nine trials. Analytic MPPI averages 22.7 km/h and completes six. Mean speeds include unsuccessful trials.

Dynamics prediction error

For each five-second segment, all three models start from the same measured state and replay the same recorded commands without resets. We compare the learned model with two analytic variants, one with fixed friction and one given the recorded contact friction.

Five-second predictions on Snow + Slope: forward and reverse paths, wrapped hitch-angle error, and longitudinal velocity error. Learned dynamics stay closer to the measured forward motion and delay reverse hitch-error growth, but large reverse errors develop later.
Five-second predictions on Snow + Slope. Forward driving is shown above, reverse below. Dotted black paths are measured trajectories; colored paths are predictions. Blue shading indicates snow.

Across the eight selected direction–environment segments, learned dynamics achieve the lowest position RMSE in six and the lowest hitch-angle RMSE in seven. Recorded friction information does not consistently close the analytic model’s prediction gap.

On Snow + Slope, the learned model delays the growth of reverse hitch-angle error, which still becomes large late in the rollout. The five-second diagnostic exceeds the controller’s planning horizons of 4.0 seconds forward and 3.25 seconds reverse.

How much history?

On a separate Snow + Slope segment, we compare one, four, and eight steps of history, plus analytic MPPI. Each predictor has three trials per direction, at targets of 30 km/h in reverse and 80 km/h forward.

Completed trials out of three
PredictorReverseForward
1 step0 / 30 / 3
4 steps (standard)3 / 30 / 3
8 steps3 / 33 / 3
Analytic MPPI2 / 30 / 3

The four- and eight-step predictors complete all reverse trials. Only the eight-step predictor completes all forward trials, averaging 64.0 km/h against an 80 km/h target.

Reverse driving across all four environments

Asphalt. Flat, uniform high-friction pavement.
Snow. Reverse driving through changing low-friction patches.
Slope. Reverse driving on asphalt with elevation changes and banking, without snow.
Snow + slope. Reverse driving with friction, grade, and banking variation.