An active-suspension system paired with an airborne jump controller passed a post-touchdown attitude test in all 24 runs at a 40° ramp and 13 m/s. The comparison controller, DART, passed in none of 24 runs. The test required pitch to stay within 35° and roll within 30° for two seconds after touchdown.
The result comes from an arXiv preprint based on simulation rather than a physical-vehicle trial. The study reports 600 new simulation runs across 72 independent BeamNG sessions.
The extra control begins before takeoff
The framework uses ramp-face suspension preconditioning to shift takeoff pitch, pitch rate and wheel spin while accounting for the control authority left for the airborne phase. Local calibration, support-aware selection and an exact-pair audit were part of the method.
Controller order was randomized within paired repetitions, BeamNG was restarted between sessions, and session-level contrasts—not individual runs—were used as the main inferential unit.
Two selected tests produced large gaps
At 11.5 m/s, a 0.35-second timing action passed the post-touchdown criterion in 23 of 24 runs. A static suspension preset passed in none of 24 runs. The comparison passed its speed-matching gate and had a Holm-adjusted p value of 0.0156.
A 200 rad/s command guard was associated with zero drivetrain hard-limit exceedances across all 600 runs.
The preset was also associated with changes in the simulated takeoff state without a meaningful shift in entry speed or flight time. Across five ramp angles at 13 m/s, the pitch bias ranged from +2.2° to +3.4° and takeoff-rate shifts from 0 to +7° per second, while entry speed and flight time stayed within 1%.
The response did not move in one direction at every speed
On a 35° tabletop fillet, median preset-minus-neutral landing-pitch changes were +13.20° at 10 m/s, +2.85° at 11 m/s, -1.35° at 11.5 m/s, -3.00° at 12 m/s and +13.25° at 13 m/s. The pattern supported a local release-phase valley at 11.5–12 m/s for the tested geometry.
On a 20° reference kicker with a 35 m lip radius, the nose-up preset retained +2.5° to +4.5° of takeoff-pitch authority. Yet at 18 m/s, median landing error worsened by 3.4° as the takeoff-rate response reversed. The reported safeguard was specific to that ramp profile.
The wheel-speed audit found a gap between the nominal mean-wheel estimate and realized per-wheel values. Median absolute error was 15.85 rad/s and maximum error was 152.28 rad/s; 60 of 240 audited strategy runs exceeded 200 rad/s on at least one wheel, although none exceeded the 221.2 rad/s hard limit.
Selection checks were largely faithful, but one interaction remained uncertain
Across 3,465 queries at 21 calibrated conditions, the selector followed all 48 prespecified positive branches, all 24 negative branches and all 24 uncertainty-boundary abstentions. The realized-interval audit accepted 42 of 48 positive selected actions.
At 35° and 16 m/s, neutral and DART each recorded 0 of 24 successes, preset-only recorded 14 of 24, and outcome-only, budget-only and joint DART-S each recorded 24 of 24. But the exact session-level interaction test was not significant: p = 0.0625.
What the simulations cannot settle
The evidence is simulation-only and local to the modeled vehicle, ramp profiles, speeds, action identities and session protocols. It does not establish performance or safety on physical vehicles, with human drivers or in field trials.
The high-speed reversal and the non-significant factorial interaction mark the boundaries of the reported evidence. The results do not show uniform improvement across speeds and ramp geometries or establish a universal high-speed safeguard.
Paper data and sources
Original title: DART-S: Reachability-Audited Active-Suspension Preconditioning for Off-Road Vehicle Jumps
Authors: Yu Hu, Fangzhou Zhao, Liang Chen et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text