← Paper Digest

Sideslip-Conditioned Drift Trajectory Planning with Flow Matching and Control Barrier Functions

Seokjun Lee, Seung-Hyun Kong

November 2026

TL;DR A flow-matching planner generates drift trajectories for a requested peak sideslip angle, control barrier functions correct them for road, friction and actuator limits, and an NMPC tracks them. In the Assetto Corsa simulator the car's median peak sideslip follows requests from 20° to 45° with a mean absolute error of 1.2°.

What they do

Existing autonomous drifting methods fix the drift intensity before deployment, through a reference trajectory, an optimization objective or an RL reward. This planner instead takes the requested peak sideslip angle as an input. A conditional flow-matching model generates drift trajectories for that request, a control barrier function (CBF) correction makes them respect road, tire-friction and actuator limits, and a nonlinear MPC tracks them.

How it works

  • Data. About 120 min (73 laps) of demonstrated drift laps on the test circuit. Each 6 s window is labeled in hindsight with its peak absolute sideslip angle, and trajectories are encoded as coefficients on smooth basis trajectories.
  • Generator. A transformer velocity field sees a 4 s measurement history, the vehicle state, 450 m of road ahead and the requested angle, and produces a plan in four Euler steps. Flow policy optimization then fine-tunes it with rewards for barrier violations, smoothness and sideslip error.
  • CBF correction. An inverse vehicle model estimates the required tire forces. Barriers cover road boundaries, axle friction utilization, speed and sideslip bounds, engine power and steering authority, and an uncertified plan is corrected in up to 8 QP steps in coefficient space.
  • Execution. 16 candidates are generated, corrected and ranked, and an NMPC with a single-track brush-tire model tracks the chosen one. Planning takes 105-126 ms (median).

Results

All runs are in Assetto Corsa (BMW M3 E92 on the Magione circuit), with 30 laps (about 75 km) per requested angle.

  • The median peak sideslip follows requests from 20° to 45° with a mean absolute error of 1.2°. Planned peaks are within 0.3° of the request, so the remaining error arises in execution.
  • Up to 35° there are no spins and at most 3 off-road events per request. At 45° there are 21 off-road events and 5 spins.
  • With the fine-tuned model, CBF correction cuts off-road events from 5.09 to 0.91 per 10 km and spins from 0.93 to 0.11. Fine-tuning lowers the sideslip error from 2.6° (supervised only) to 1.2°.

Limitations

Results come from simulation only; real-vehicle experiments are future work. The demonstrations come from the same circuit used for testing. At a 20° request the car’s median peak is 16.4°, and road departures rise at 40-45°. The NMPC model does not capture tire temperature, pressure or fuel-mass changes, which the authors name as a source of tracking error.

Videos

Full video (YouTube)

Figures

From Figure 2: Overview of the planner. Flow matching generates candidate trajectories, the CBF correction addresses ...
From Figure 2: Overview of the planner. Flow matching generates candidate trajectories, the CBF correction addresses the vehicle constraints, and the NMPC tracks the selected trajectory.
From Figure 3: Closed-loop drift trajectories and sideslip profiles at two corners of Magione for six requested peak ...
From Figure 3: Closed-loop drift trajectories and sideslip profiles at two corners of Magione for six requested peak sideslip angles. Vehicle poses are colored by sideslip magnitude, and dashed lines indicate the requested angles.
From Figure 4: Closed-loop drifting in Assetto Corsa for six requested peak sideslip angles. Each panel shows the cha...
From Figure 4: Closed-loop drifting in Assetto Corsa for six requested peak sideslip angles. Each panel shows the chase camera and the selected trajectory.
Abstract

Drifting, the deliberate control of a vehicle at large sideslip angles, is relevant to both motorsport and recovery from unexpected oversteer. Existing autonomous drifting methods typically determine drift intensity through predefined reference trajectories, optimization objectives, or learned policy rewards, rather than accepting a desired sideslip angle as an explicit input. We propose, to the best of our knowledge, the first generative drift trajectory planner explicitly conditioned on a requested peak sideslip angle. A conditional flow matching model is trained on demonstrated drift laps with hindsight-derived sideslip labels and refined through reinforcement fine-tuning. A control barrier function (CBF) correction is then applied to satisfy model-based road, tire friction, and actuator constraints at the planning waypoints. In closed-loop experiments in the racing simulator Assetto Corsa, a nonlinear model predictive controller tracks the generated plans for six requested sideslip angles from 20 deg to 45 deg, with 30 laps (approximately 75 km) per request. The median peak sideslip angle measured across drift-corner passes follows the requested angle, with a mean absolute error of 1.2 deg across the six conditions. Ablation experiments show that reinforcement fine-tuning improves conditioning accuracy, while CBF correction substantially reduces off-road and spin events.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.