Beijing Jiaotong's H-HTP Hits 97% Success Rate on Highway Collision Avoidance
A hybrid planner hands lateral safety to a mixed-integer optimizer while learning only the velocity profile, hitting 97% success at 54 ms per cycle.
ProTakeaways
- H-HTP splits highway planning: a learned model predicts velocity, an MIQP solver handles all lateral and safety decisions.
- Achieves 97%+ scenario success rate with a 54 ms mean planning cycle on the HighD dataset.
- A piecewise-linear vehicle geometry cuts integer variables from Q×N to just N, one sign bit per time step.
- Beats a VectorNet-only baseline that failed to react in frontal cut-in tests, colliding with the intruding vehicle.
- Trajectory-level displacement error drops from 0.66 m (learning only) to 0.06 m with the hybrid approach.
- Full source code available at github.com/alasjia/H-HTP-Framework, built on Gurobi 11 and PyTorch.
Highway autonomy has a nasty asymmetry: catastrophic events are rare, but when a driver ahead cuts in at 30 m/s the reaction window shrinks to under a second. That combination has made trajectory planning a stubborn engineering problem. Learning-based planners struggle on the long tail, and optimization-based planners struggle to hit real-time budgets. A new framework called H-HTP proposes a clean split between the two, and backs it up with numbers from over half a million planning cycles on real German highway data.
The paper, Reliable and Real-Time Highway Trajectory Planning via Hybrid Learning-Optimization Frameworks, comes from a team at Beijing Jiaotong University, Hainan University, and UT Compiègne. The source code sits on GitHub, which is unusual for MIQP-heavy planning work and makes the results easier to reproduce.
Why hybrid planners keep tripping over themselves
Trajectory planning approaches fall into two camps. Learning-based methods achieve strong generalization through data-driven modeling but struggle to provide reliable safety guarantees in rare yet critical situations. Optimization-based methods offer formal constraint modeling and safety guarantees, but face computational challenges from the nonlinearity introduced by precise vehicle geometry and obstacle avoidance. Hybrid architectures try to get the best of both, and usually inherit the worst problems from each side.
The authors argue the failure mode is architectural. When learning modules are tasked with high-dimensional lateral decisions such as lane-change initiation or full trajectory generation, their training may fail to converge reliably in the critical cases, because dangerous highway events are so rare in real datasets. Meanwhile, mixed-integer formulations that could enforce hard safety constraints tend to blow up computationally, because the integer variable count scales unfavorably with the planning horizon.