HPTune: Hierarchical Proactive Tuning for Collision-Free Model Predictive Control

Parameter tuning is a powerful approach to enhance adaptability in model predictive control (MPC) motion planners. However, existing methods typically operate in a myopic fashion that only evaluates executed actions, leading to inefficient parameter updates due to the sparsity of failure events (e.g., obstacle nearness or collision). To cope with this issue, we propose to extend evaluation from executed to non-executed actions, yielding a hierarchical proactive tuning (HPTune) framework that combines both a fast-level tuning and a slow-level tuning. The fast one adopts risk indicators of predictive closing speed and predictive proximity distance, and the slow one leverages an extended evaluation loss for closed-loop backpropagation. Additionally, we integrate HPTune with the Doppler LiDAR that provides obstacle velocities apart from position-only measurements for enhanced motion predictions, thus facilitating the implementation of HPTune. Extensive experiments on high-fidelity simulator demonstrate that HPTune achieves efficient MPC tuning and outperforms various baseline schemes in complex environments. It is found that HPTune enables situation-tailored motion planning by formulating a safe, agile collision avoidance strategy.

💡 Research Summary

The paper addresses a fundamental limitation of existing model predictive control (MPC) auto‑tuning approaches: they rely solely on the outcomes of actions that have already been executed. Because failure events such as near‑misses or collisions are sparse, parameter updates occur infrequently, leading to sub‑optimal performance in dynamic environments. To overcome this, the authors propose HPTune, a Hierarchical Proactive Tuning framework that extends evaluation to non‑executed actions predicted by the MPC horizon. HPTune consists of two complementary layers.

Fast‑level tuning operates at a high frequency (every planning cycle) and updates a subset of parameters using lightweight risk indicators. Two indicators are defined: (1) Predictive proximity distance – the minimum distance between the ego vehicle’s predicted bounding box and each obstacle’s predicted bounding box; (2) Predictive closing speed – the component of the relative velocity projected onto the line joining ego and obstacle positions. The ratio of closing speed to proximity distance quantifies imminent collision risk. This ratio is passed through a ReLU (to ignore receding obstacles) and a tanh scaling (to keep updates smooth) to produce a proactive safety margin ϕ⁽ᵖʳᵒᵃᶜᵗⁱᵛᵉ⁾. The final safety distance used in the MPC constraints becomes ϕ_base + ϕ⁽ᵖʳᵒᵃᶜᵗⁱᵛᵉ⁾, allowing the planner to anticipate hazards before they materialize.

Slow‑level tuning runs at a lower frequency (every T steps) and performs full back‑propagation on an extended closed‑loop loss that combines three components: (i) a navigation loss L₁ that penalizes deviation from reference states and control effort over the past T executed steps; (ii) a predictive safety loss L₂ that penalizes any predicted non‑executed action whose proximity distance falls below the proactive safety margin; and (iii) an ℓ₁ regularization L₃ that discourages overly conservative safety margins. The total loss L = η₁L₁ + η₂L₂ + η₃L₃ is differentiated with respect to the tunable parameters Γ = {α, β} and updated via gradient descent. α balances tracking versus control effort, while β controls the sensitivity of the proactive safety margin.

A key technical enabler is the integration of Doppler LiDAR with a Kalman filter. Conventional 3‑D LiDAR provides only position measurements, which are insufficient for accurate closing‑speed estimation. Doppler LiDAR supplies the line‑of‑sight component of each obstacle’s velocity. The authors aggregate per‑point Doppler readings inside each obstacle’s bounding box, project them onto the obstacle’s motion direction, and feed the resulting velocity estimates into a Kalman filter to predict future obstacle states over the truncated horizon H − t. These predictions feed directly into both the fast‑level risk indicators and the slow‑level loss terms.

The framework is evaluated in the high‑fidelity CARLA simulator using a 40 m × 40 m test field populated with randomly moving 4‑wheel obstacle vehicles. Four methods are compared: HPTune (proposed), DiffTune‑MPC (closed‑loop tuning), RDA (open‑loop tuning), and OBCA (no tuning). Hyper‑parameters are set to H = 15, Δt = 0.12 s, ϕ_base = 0.2 m, ϕ_max = 2.0 m, η₁ = η₂ = 10⁻¹, η₃ = 10⁻³, T = 5, ε = 10⁻³, α = 0.2, β = 5.0.

Results demonstrate that HPTune consistently outperforms the baselines across multiple metrics. In scenarios with 4 and 6 obstacles, the pass rate improves by 8.3 %–46 % relative to the other methods, with a notable 13.7 % gain over DiffTune‑MPC in the 4‑obstacle case. Motion smoothness is also superior: average acceleration drops to 1.994 m/s² and average jerk to 5.975 m/s³, both lower than any baseline. Average navigation time is reduced to 13.732 s, again the best among the competitors. A runtime probability density function of safety distances shows that RDA adjusts safety only when obstacles are already close, causing abrupt control spikes, whereas HPTune continuously adapts the safety margin throughout the trajectory, yielding smoother throttle and steering profiles. Moreover, the pass‑rate curve converges after roughly 200 parameter‑update iterations, indicating rapid tuning efficiency.

The authors conclude that extending closed‑loop evaluation to predicted, non‑executed actions transforms MPC tuning from a reactive, failure‑triggered process into a proactive, anticipation‑driven one. By leveraging Doppler‑enhanced obstacle predictions, HPTune can adapt safety margins on the fly, delivering collision‑free, agile navigation even in densely populated, dynamic environments. Future work is suggested on real‑time implementation of Doppler processing, computational optimization of the Kalman filter, and validation with heterogeneous agents such as pedestrians and cyclists.

Overall, HPTune represents a significant step toward more resilient autonomous vehicle control, offering a practical recipe for integrating predictive perception data into hierarchical parameter adaptation mechanisms.