A corridor where nobody has validated a merge policy for articulated trucks needs one built and validated for it. Represent the trailer in the planner with an articulated hitch-angle kinematic model. Check its envelope with swept-path and offtracking analysis. Switch to a dynamic trailer model for high-speed, high-lateral-acceleration manoeuvres. Take gap thresholds from measured truck lane-change data, not from car merge studies. Each threshold then needs its own argument in the safety case, because UL 4600 requires a safety argument but sets no pass/fail criteria.
The specifics here are for tractor-trailers merging from highway on-ramps.
Which kinematic model should represent trailer tail swing in the planner?
Start with an articulated kinematic model that carries a hitch-angle state. It adds the trailer heading, or the relative angle between trailer and tractor, to the tractor’s state. The geometry comes from three lengths: the tractor wheelbase, the distance from the rear tractor axle to the hitch, and the distance from the trailer axle to the hitch. A proof-of-concept truck-parking study concludes that until the two-body dynamics are represented correctly, higher planning and control layers cannot be meaningfully integrated or evaluated. Hybrid A* suits articulated vehicles because it builds the nonholonomic constraints into the search itself, including minimum turning radius and maximum hitch angle.
The hitch angle has to be a hard constraint at all times, because exceeding the limit while reversing puts the vehicle into a jackknife. The same study still saw jackknifing in some configurations, and its authors say their hitch-angle enforcement was not yet robust across all manoeuvres. A hitch-angle constraint written into the planner still has to be tested. It is not a guarantee.
In MERL’s 2023 planner for semi-trailers, the trailer appears in the planner’s cost, not only in its collision check: the lateral deviation of the trailer’s rear-axle point is computed from the hitch angle, and the tractor’s and the trailer’s deviation from lane centre are both penalised. MERL’s paper also accepts that the hitch angle is only estimated online, with finite precision, and handles this by putting uncertainty into the vehicle model. In closed-loop simulation, tracking error stayed under 10 cm across its scenarios, which included lane following, lane change and obstacle avoidance.
The no-slip assumption caps the model’s useful speed
The kinematic trailer model assumes the tyres do not slip sideways. That holds at the low speeds where parking happens, and it is why kinematic models are most accurate there. A Linköping University thesis on tractor-trailer control found that controllers built on kinematic models perform well up to about 55 km/h, after which tractor error climbs quickly. In simulation, both kinematic controllers failed above 90 km/h because they could not handle the trailer’s momentum, and the vehicle rolled over. The same thesis found that adding a trailer model cut position prediction error by around a third, and concluded that a model of trailer dynamics is necessary for safe control at high speed.
That matters for merging, since one planner’s on-ramp test cases spanned main-lane speeds of 10 to 120 km/h. We would keep the kinematic model for the search, where it is cheap, and use a dynamic trailer model to predict and check collisions at merge speed.
Swept path and offtracking as the design check
Offtracking is the difference between the paths of the rear wheels and the front wheels in a turn. It generally grows with axle spacing and shrinks on larger-radius turns. For geometric design, the Federal Highway Administration (FHWA) uses swept path width: the distance between the outside front tractor tyre and the inside rear trailer tyre. Low-speed offtracking depends mainly on the distance from the kingpin to the centre of the trailer’s rear axle group, which is the semitrailer’s effective wheelbase. At higher speeds, lateral acceleration pushes the rear outward until, above a certain speed, the rear axles track outside the tractor steering axle. On sweeping high-speed curves this is 1 foot or less at 55 mph for most configurations FHWA analysed.
AASHTO design vehicles are a useful outer check, because access and exit ramps are designed to accommodate them. A truck longer than the design vehicle will run its trailer wheels off the pavement on the inside of a curve.
Tail swing needs checking separately from cut-in. In instrumented tests of a 16 m steered semitrailer, the steering system cut maximum cut-in by about 67% but substantially increased tail swing, and it raised rearward amplification by 20%. Tail swing grew because shortening the effective length increases the effective rear overhang. Even on a conventional rig, hitch placement relative to the tractor’s rear axle causes yaw-induced swing that a simple model will not match exactly. An envelope tuned to reduce cut-in can make a lane change worse.
Does the kinematic model need re-deriving for each trailer length and articulation point we haul?
The equations do not need re-deriving, but the parameters do. The kinematic structure generalises: there is a recursive formula for an n-trailer kinematic model. What changes between configurations includes the kingpin setting and rear overhang, which drive the envelope. FHWA notes that a 53-foot semitrailer with a 41-foot kingpin setting offtracks the same as a 48-foot one, yet its rear swings out further because of the extra five feet behind the axle group. A longer kingpin setting also means a larger turning radius. US federal law sets no requirement for kingpin setting, kingpin setback or rear overhang, with 41 feet the most common kingpin setting. Where the hitch sits relative to the tractor’s rear axle also affects how well the tractor can steer the trailer into the orientation the planner wants. Load is a parameter too, because high-speed offtracking depends on axle loads and suspension.
In practice, re-identify parameters for each configuration from measured data. The Linköping work identified parameters by nonlinear optimisation, assuming perfect measurement of all states, minimising the one-step prediction error. Measurable quantities such as axle positions and centre of gravity were allowed to drift from their measured values where that improved the fit. The models were then validated over a 1 s prediction horizon, against a 10 ms identification horizon, on double lane changes at varying speeds, scored by RMSE.
A 2003 instrumented trailer-steering test gives a pattern for on-road validation. An instrumented articulated vehicle ran a 90-degree corner, a steady-state circle and a lane change at low and high speed, and the measurements were compared against a yaw-roll model. Simulated lateral tyre forces compared favourably with the tests. With the trailer steering free, they were slightly overestimated but still acceptably accurate.
Can you adopt an existing trailer kinematics library instead? The parking study simply adapted an open-source A* planning simulation to carry a tractor-trailer model. Whether you build or adopt, each configuration still needs its own identification and validation.
| Layer | Model | Where it holds | Validate against |
|---|---|---|---|
| Planner search | Articulated bicycle / hitch-angle kinematic model with a hard hitch-angle limit | Low speed; tractor error grows above about 55 km/h in the Linköping simulations | Closed-loop tracking in simulation, then measured runs |
| Prediction and collision check at merge speed | Dynamic trailer model with tyre slip | High speed and large lateral accelerations | Double lane changes at varying speeds, scored over a 1 s horizon |
| Envelope design check | Swept-path width and offtracking | Geometric design of ramps and turns | Measured cut-in, tail swing and overshoot, as in the steered-trailer tests |
How big a lane-change gap does a long trailer need?
The defensible starting point is measured truck behaviour, adjusted for each trailer’s geometry and checked with swept-volume collision analysis. The best truck-specific data we found is an analysis of 600 highway lane changes drawn from the University of Michigan Transportation Research Institute’s IVBSS naturalistic database, which covers 10 Class 8 tractors. Thirteen left lane changes involved a time-to-collision under 4 s with required deceleration above 0.8 m/s², while the largest required deceleration seen in right lane changes was about 0.33 m/s². The authors propose warning conditions of TTC under 4 s or required deceleration above 0.8 m/s² when the truck is slower than the vehicle in the target lane. For right lane changes with the truck faster, they propose a range threshold of 12.7 m.
How the gap is measured matters for a trailer. Both IVBSS analyses measure the rear gap from the rear edge of the subject vehicle’s trailer to the front of the following vehicle, not from the tractor.
On-ramp merges are mandatory lane changes, which are more aggressive than discretionary ones. In a companion study of 640 mandatory and 2,035 discretionary truck lane changes, mandatory changes were about 4.5 times more likely to involve a critical gap and spent over 10% less time crossing the lane line. A threshold set from mostly discretionary manoeuvres can therefore understate the pressure the planner faces in a mandatory merge.
| Measured threshold (IVBSS truck data) | Value | Condition |
|---|---|---|
| Time-to-collision | under 4 s flags risk | Truck slower than target-lane vehicle |
| Required deceleration | above 0.8 m/s² flags risk | Truck slower than target-lane vehicle |
| Rear range | under 12.7 m flags risk | Right lane change, truck faster |
Know the data’s limits before you lean on it. It comes from one fleet with 18 male drivers, trucks governed to 62 mph, and events screened for straight roads at 55 mph or above. Car merge studies are no substitute: one widely cited on-ramp study removed trucks from its dataset altogether, and it notes that drivers often merge into gaps smaller than the critical gap. A 2022 review of on-ramp merging control adds that strategies for mixed traffic tend to underestimate human-driver uncertainty and overestimate drivers’ willingness to cooperate.
One conservative rule for 16 to 34 ft, or thresholds per length bucket?
Treat this as an empirical question with a clear test. Run the same scenario set across the whole range with a single conservative threshold and measure two things: safety margin and missed merges. If one rule keeps both acceptable at every length, adopt it. If it does not, define buckets and validate each one separately with swept-path and collision analysis. If you do need buckets, define them by effective wheelbase, hitch offset and rear overhang rather than by nominal trailer length. FHWA’s comparison shows that two trailers of different length can offtrack identically and still swing differently at the rear. Two caveats remain: high-speed offtracking also depends on speed and axle loads, and a policy validated on one corridor’s geometry may not carry over to another on-ramp.
Building a merge-gap policy for oversized trailers on one corridor’s on-ramps
The merge-gap policy has to be built and validated for the corridor’s own ramp geometry and trailers before the pilot.
TechnoLynx frames the requirement in terms of the corridor’s operating domain, ramp geometry and trailer dynamics, then encodes, tunes and regression-tests the policy as part of its AI engineering for manufacturing and automotive work, delivering a merge-gap policy validated for the corridor’s safety-case sign-off. TechnoLynx also supports rollout and monitors live merge performance against the safety case.
The SMART-Merge planner (arXiv, 2025) was verified on 160 test cases built from eight US on-ramps, with main-lane speeds from 10 to 120 km/h, designed to reproduce traffic seen in naturalistic data. A merge counted as a success only if it finished with no collision, no contact with the road boundary, and within a 100 s timeout. Reuse those success criteria and the way the cases were generated. The ego model does not transfer: SMART-Merge uses the classical kinematic bicycle model of a single rigid vehicle, so a trailer version has to swap in the articulated model and measure following distances to the trailer’s rear edge. One finding does transfer directly. Desired speed should come from safety requirements within the merge window rather than from the speed limit, and in SMART-Merge’s tests the version that simply targeted the speed limit merged successfully far less often.
Encoding ramp curvature and merge-lane length
The merge must be completed before the end of the acceleration lane. In car merging data, a slow-down pattern became likelier nearer the end of the acceleration lane, and distance from the ramp entrance was one factor shaping merging behaviour. For lateral control on a curved ramp, a Frenet frame is natural: position becomes progress along the path plus signed lateral distance from it. Some trucks roll over at lateral accelerations as low as 0.30 g, while AASHTO policy allows up to 0.17 g, which leaves little margin, so ramp design speed needs care.
The validation sequence
- Specify the corridor’s operating domain: the ramps, speeds, traffic and trailer configurations the policy must handle.
- Model trailer dynamics and corridor geometry in a vehicle-dynamics simulator such as TruckSim or an extended CARLA, with each trailer configuration’s identified parameters.
- Explore gap thresholds in simulation, then calibrate them against recorded real-world merges on similar ramps.
- Run closed-loop regression at fleet scale, including a scenario library that stresses tail-swing cases beyond those actually driven.
- Take the policy, its evidence and its known limits through hazard analysis and safety-case sign-off before pilot deployment.
Simulated trailer swing must be validated against the real vehicle, or the thresholds may not transfer. Thresholds tuned to one corridor may not generalise to the next.
What UL 4600 decides and what it leaves to the corridor team
Underwriters Laboratories published UL 4600 in April 2020. It requires a safety argument covering risk analysis, testing, autonomy validation and data integrity, among other topics. It is technology neutral, defines no performance or pass/fail criteria, and does not cover road testing or acceptable risk levels. Koopman and Laplante (IEEE Computer, 2023) describe it as an umbrella that coordinates software and system safety standards so nothing is left out. Our guide to constructing a defensible safety case for autonomous system sign-off covers how that argument is structured.
So the standard will not supply a TTC figure or a gap length for your merge policy. Each number has to be justified from your own evidence: the identified parameters for each trailer, the measured truck gap data and its limits, and the regression results on the corridor’s geometry. Explicit numerical limits do exist in other approval regimes. FHWA describes an approval for 88,000-pound A-train milk trucks that required a static roll threshold of at least 0.45 g, a dynamic load transfer ratio of at most 0.6, and high-speed transient offtracking of at most 0.5 m. Those limits come from a performance-standards approval for milk trucks, not an autonomous merge. That approval sets each limit and requires it to be met under all loading conditions. A corridor policy that leaves out either is not finished, however well it performs in simulation.
FAQ
Which type of truck and trailer combination is most likely to have a crack-the-whip rollover?
According to FHWA, trailers in double-trailer combinations roll over more often than single trailers. The whip effect is measured as rearward amplification: the rear trailer’s peak lateral acceleration divided by the power unit’s in a rapid steering manoeuvre.
What is a tractor-trailer combination?
A tractor-trailer is an articulated vehicle: a tractor and a trailer joined by a hitch, which adds degrees of freedom that have to be actively managed. On a semitrailer, the pivot is the kingpin in the fifth-wheel connection.
