Source-linked AI summary
A Differentiable Chassis Safety Layer for Intelligent Vehicle Motion Control
Lu Xiong, Xuge Hu, Yufei Song, Weidong Liu, Chen Tang
TL;DR
Vehicle motion control must coordinate task objectives with state-dependent chassis constraints, but independent safety projections may compromise task performance. ChassisOptNet uses sequential dynamics-coupled QPs and two-stage learning to coordinate policy decisions with safety projection. Closed-loop evaluations report improved task execution under evaluated generalization and emergency conditions.
Problem
Vehicle motion control must coordinate diverse objectives with coupled chassis constraints that vary with operating conditions, while independent projections may modify task commands without accounting for performance degradation.
Source: §1 · pp. 1–2Method
ChassisOptNet embeds actuator limits, individual-tire friction constraints, and handling envelopes in sequential QPs coupled through predicted vehicle dynamics, with two-stage expert-imitation and risk-guided learning.
Source: §1 · p. 2, §3.3.2 · p. 9Results
Closed-loop CarSim experiments demonstrate improved task execution under the evaluated generalization and emergency conditions.
Source: §6 · pp. 13–14Takeaways & Limitations
The safety layer provides a modular interface for task-oriented control and can extend to other vehicle motion-control objectives with appropriate task formulations.
Source: §5 · p. 13The simplified vehicle model, relaxed safety constraints, and assumed state availability preclude strict closed-loop safety guarantees.
Source: §5 · p. 13
Abstract
from arXiv · showhide
Learning-based vehicle motion control requires coordination between diverse task objectives and state-dependent chassis safety constraints. Conventional cascaded architectures may compromise task performance when safety filters modify nominal commands independently of policy learning. This article proposes a differentiable chassis safety layer that modularly integrates task-oriented control with explicit physical constraints. Actuator limits, individual-tire force constraints, and vehicle handling envelopes are incorporated into sequential quadratic programs. Each projected command propagates the predicted vehicle dynamics and updates subsequent constraints, enabling temporal gradient coupling through implicit differentiation. A two-stage learning strategy combines expert imitation with risk-guided supervision and conditional intervention guidance to coordinate nominal control decisions with chassis safety constraints. Closed-loop evaluations under unseen scenarios, emergency maneuvers, and ablation studies demonstrate that the proposed method improves safety-performance coordination and exploits vehicle dynamic potential while maintaining handling stability.
1 Introduction
Vehicle motion control must coordinate multiple task objectives with chassis constraints that vary with vehicle state, while independent safety projections can modify task commands without accounting for performance degradation. ChassisOptNet integrates task learning with chassis constraints through sequential, dynamics-coupled QPs and risk-guided training.
1 Introduction
- Independent constraint projections may modify task-oriented commands without considering resulting performance degradation, while optimization-layer gradients can become uninformative under some active-constraint configurations.MPC effectiveness also depends on prediction accuracy, particularly with nonlinear tire behavior, uncertain adhesion, and model mismatch.Source: §1 · pp. 1–2
- Chassis constraints are dynamically coupled: tire-force capacity depends on force demands, vertical loads, and road adhesion, while handling limits depend on evolving sideslip and yaw.Each constrained command changes subsequent vehicle states and constraint conditions, a temporal interaction independent single-step projections cannot explicitly represent.Source: §1 · p. 2
- ChassisOptNet integrates actuator limits, individual-tire friction constraints, and handling envelopes into state-dependent QP layers for joint task–constraint learning.The framework is demonstrated on four-wheel-steering vehicle motion control, using trajectory tracking as a representative task.Source: §1 · p. 2
- Sequential QPs propagate each projected control action through predicted vehicle dynamics, updating later states and constraints while providing temporal sensitivities for upstream learning.The framework uses implicit differentiation to obtain these projection sensitivities.Source: §1 · p. 2
- A two-stage learning strategy combines expert imitation with expert-relative risk supervision and conditional safety-intervention guidance to coordinate task control with chassis projection.Closed-loop simulations validate the framework under unseen and near-limit operating conditions.Source: §1 · p. 2
2 Differentiable Chassis Safety Layer
The safety layer maps chassis limits into control-space constraints and sequential differentiable projections. Vehicle-state propagation links those projections so later constraints and their sensitivities can inform policy learning.
2 Differentiable Chassis Safety Layer
- Steering-actuator, tire-force, and handling-stability limits are unified in control space for sequential safety projections.Each projected steering command updates the predicted state and subsequent constraints, enabling temporal gradient propagation during policy training.Source: §2 · p. 3
2.1 Vehicle Dynamics for Safety Projection
- A two-degree-of-freedom bicycle model with linear tire cornering stiffness predicts lateral and yaw motion for the safety layer.Longitudinal speed is treated as an available parameter; nonlinear tire saturation is addressed through additional friction constraints.Source: §2.1 · p. 3
2.2 Multi-Level Chassis Constraints
- Three constraint classes—control-space, state–control coupled, and state-space—form a unified control-space representation for safety projection.They cover actuator limits, tire friction mapped through the tire model, and handling-envelope limits converted using control barrier functions.Source: §2.2 · p. 3
2.2.1 Constraints in Control Space
- Steering-angle bounds constrain commands directly, while steering-rate limits apply between successive projected commands.At the first prediction step, the rate bound references the previously applied steering command; later steps use the preceding projected command.Source: §2.2.1 · p. 3
2.2.2 Constraints in State–Control Coupled Space
- Per-wheel tire-force constraints map state-dependent lateral-force capacity into steering inequalities while retaining individual wheel limits and slack penalties.A negative unmodified friction radicand means longitudinal-force demand already exceeds reserved friction capacity; the numerical lower bound does not restore feasibility.Source: §2.2.2 · p. 4
2.2.3 Constraints in State Space
- A four-sided lateral–yaw handling envelope is converted into steering-dependent inequalities using relative-degree-one control barrier functions.Forward invariance is established for the nominal continuous-time model only when the CBF inequalities are feasible and enforced without relaxation; the soft-constrained filter does not establish strict invariance.Source: §2.2.3 · pp. 4–5
2.3 Chassis-Constrained QP Formulation
- Over a horizon matching the steering-policy output horizon, sequential QPs minimize local steering corrections and propagate projected commands through predicted vehicle states.The first projected command is applied during receding-horizon safety-filtered control; remaining projected commands and states provide temporal information for offline policy training.Source: §2.3 · pp. 5–6
2.4 Dynamics-Unrolled Differentiable Projection
- A differentiable QP layer parameterizes projection weights, CBF gains, and friction reserve to adapt the admissible control region under varying conditions.Under appropriate regularity conditions, differentiating the KKT conditions yields QP sensitivities for learning projection parameters and coordinating policy with safety.Source: §2.4 · p. 6
3 Safety-Guarded Task Learning
The framework coordinates trajectory tracking with physical constraints through a differentiable handling-stability projection and a two-stage learning strategy. Sequential projection sensitivities connect task-oriented learning to the effects of control corrections on later states and constraints.
3.1 Modular Architecture and Signal Flow
- A policy network, safety-parameter generation network, and differentiable handling-stability projection layer coordinate nominal steering with physical constraints.Both networks process vehicle observations through a shared feature representation.Source: §3 · p. 7, §3.1 · p. 7
3 Safety-Guarded Task Learning
- The framework combines nominal steering prediction, safety-parameter generation, and differentiable projection to coordinate tracking performance with physical constraints.Task-related losses and projection sensitivities provide upstream gradient feedback, depending on enabled losses and trainable parameter groups.Source: §3 · p. 7, §3.1 · p. 7
3.2 Projection-Guided Gradient Coupling
- Sequential QP Jacobians capture local sensitivities and how earlier control corrections affect later vehicle states and constraints, coupling task learning with chassis constraint projection.Projection weights set the control-correction metric, while CBF gains and tire-friction margins parameterize constraints.Source: §3.2 · p. 8
3.3 Two-Stage Task–Constraint Learning
- Training first initializes the nominal steering policy from expert MPC demonstrations, then adds handling-risk supervision and differentiable safety projection.Stage II is intended to improve coordination between tracking performance and physical constraints.Source: §3.3 · p. 8
3.3.1 Task-Oriented Learning
- Stage I trains the nominal policy to predict expert steering sequences with imitation and steering-smoothness losses, without activating the safety projection.The sequence objective applies prediction-step weighting.Source: §3.3.1 · p. 8
3.3.2 Risk-Guided Fine-Tuning
- Stage II combines expert imitation, risk-gated intervention, expert-relative risk supervision, curvature-dependent tracking, and QP slack regularization.Risk supervision penalizes excess risk, boundary violations beyond budget, peak risk, and unfavorable risk growth relative to fixed expert trajectories; it does not impose a terminal invariant-set condition.Source: §3.3.2 · pp. 8, 10
- Risk-gated intervention losses guide the nominal policy toward feasible steering when predicted risk exceeds the expert-relative allowance.The intervention gate and projected steering targets are detached during differentiation.Source: §3.3.2 · p. 9
- Stage II updates the policy and shared features, while the safety-parameter output layer is optimized only when enabled and remains fixed in the reference configuration.Projected-sequence imitation and optional nominal-rollout tracking have zero loss weights in that configuration.Source: §3.3.2 · p. 10
4 Experimental Validation
Closed-loop Python–CarSim simulations evaluated generalization, near-limit handling, and emergency maneuvers, with comparisons against model-based controllers and ablations. The results report improved tracking in several tested conditions, while also identifying transient handling-boundary excursions.
4 Experimental Validation
- Closed-loop simulations assessed varying trajectories, speeds, and road adhesion, a near-limit maneuver, and ablations of safety projection and learning.Generalization tests used held-out conditions, including speeds of 65–80 km/h and adhesion coefficients of 0.3–0.5.Source: §4 · p. 10, §4.1 · p. 12, §4.3 · p. 13
4.1 Simulation Platform and Expert Demonstrations
- The expert dataset comprised 181 successful trajectories across seven path families, with reference speeds of 30–60 km/h and adhesion coefficients of 0.6–1.0.Trajectories were split 75:25 into training and validation subsets; the policy was trained first by imitation and then by safety-guided fine-tuning.Source: §4.1 · p. 11
4.5 Ablation Study of Safety Projection and Learning
- 0.3422 m maximum and 0.1011 m mean lateral errors were achieved by the complete framework, versus 0.4562 m and 0.1150 m for the nominal policy in emergency maneuvers.The fixed non-differentiable filter failed to complete the maneuvers, with maximum lateral errors exceeding 3 m.Source: Table 4 · p. 12, §4.5 · p. 13
4.2 Baselines and Evaluation Metrics
- LQR served as an unconstrained linear-feedback reference and MPC as the model-based tracking benchmark; evaluation included tracking errors and lateral–yaw handling responses.Larger sideslip and yaw-rate magnitudes alone do not establish unsafe behavior; interpretation also requires handling boundaries and tire-force limits.Source: §4.2 · pp. 12–13
4.3 Generalization to Unseen Operating Conditions
- Mean lateral tracking error fell from 0.0358 to 0.0199 m in trajectory generalization and from 0.0449 to 0.0231 m in adhesion generalization versus MPC.Under velocity generalization, the proposed method had a lower maximum error of 0.1783 m, but its 0.0443 m mean error was slightly higher than LQR's.Source: Table 3 · p. 11, §4.3 · p. 13
4.4 Near-Limit Handling Assessment
- In a low-adhesion J-turn, the proposed controller converged faster than MPC but transiently exceeded the prescribed handling boundary.LQR showed substantial trajectory deviation and unstable lateral–yaw responses; MPC mitigated instability but converged more slowly.Source: Figure 3 · p. 12, §4.4 · p. 13
5 Discussion
The discussion presents the safety layer as a modular interface and reports benefits from coordinating policy learning with chassis safety projection, while noting limits to safety guarantees.
5 Discussion
- The architecture may extend beyond trajectory tracking to other vehicle motion-control objectives when paired with appropriate task formulations.The reported generalization and near-limit experiments show improved tracking under evaluated conditions, while ablations highlight benefits from coordinating learning and projection.Source: §5 · p. 13
- The simplified vehicle model, relaxed safety constraints, and assumed state availability preclude strict closed-loop safety guarantees.Projection-gradient effectiveness also depends on training configuration; future work includes uncertainty-aware projection, real-time implementation, and experimental validation.Source: §5 · p. 13
6 Conclusion
The conclusion describes a differentiable chassis safety layer that integrates physical constraints into sequential QPs and supports safety-guided learning. Closed-loop CarSim experiments report improved task execution under evaluated generalization and emergency conditions.
6 Conclusion
- Actuator limits, individual-tire friction constraints, and vehicle handling envelopes are embedded in sequential QPs coupled through predictive vehicle dynamics.The layer provides projection sensitivities for safety-guided learning, alongside a two-stage strategy combining expert imitation, risk supervision, and conditional intervention guidance.Source: §6 · p. 13
- Closed-loop CarSim experiments show improved task execution under the evaluated generalization and emergency conditions.The conclusion says these results support the effectiveness of the task–safety coordination framework.Source: §6 · p. 14
Author Biographies
The biographies describe six authors’ education, appointments, and research interests in vehicle engineering, intelligent vehicles, and related areas.
Author Biographies
- Lu Xiong is a Tongji University professor and researches intelligent-vehicle perception, planning, control, estimation, and testing.He received a vehicle-engineering Ph.D. from Tongji University in 2005.Source: Author Biographies · p. 15
- Xuge Hu is pursuing a Ph.D. at Tongji University and researches intelligent vehicle control and autonomous driving.She received a transportation B.S. from Jilin University in 2024.Source: Author Biographies · p. 15
- Yufei Song’s stated research interests are autonomous driving and motion planning.He received B.S. and M.S. degrees from Tongji University in 2023 and 2026, respectively.Source: Author Biographies · p. 15
- Weidong Liu is a doctoral student at Tongji University and works at Jiangling Motors; his interests include autonomous-vehicle decision, planning, and dynamic control.He received an M.E. degree from China Agriculture University in 2012.Source: Author Biographies · p. 16
- Chen Tang is an associate professor at Tongji University whose research interests include advanced chassis systems, active vehicle safety, and intelligent transportation.He received a mechanical and mechatronics engineering Ph.D. from the University of Waterloo in 2018.Source: Author Biographies · p. 16