Training-Independent Validation of a Learning-Based Controller for Stunt Maneuvers
Authors: Kiruthiga Chandra Shekar, Aliasghar Arab, Jingang Yi
Organizations: General Autonomy, USA · City College of New York, New York University, and General Autonomy, USA · Department of Mechanical and Aerospace Engineering, Rutgers University, Piscataway, NJ, USA
Learned controllers for vehicle maneuvers are usually validated under the conditions and objective used to train them, where they appear highly reliable, yet the parameters that change on real roads are often not those the model was trained on. We argue that validation must be independent of training: it should vary physical parameters that training kept fixed and use acceptance criteria absent from the training objective. A learning-based control policy, G-Maneuver, with 90.11% success is tested for stunt-level maneuvers. A looser terminal-heading score ranked the worst setting as the best. Because real stunt tests are safety-critical, a simulation-to-reality-to-simulation (Sim2Real2Sim) loop uses the simulation to select safe zones for real-vehicle trials with 16.9% acceptance. Automotive safety analysis frameworks are used to identify parameters for validation using coverage tests. We demonstrated that the kinematic stunt-driving indicators, namely heading change, sideslip, and speed-normalized yaw rate, match between simulation and the real world, while force and roll indicators do not, which identifies the indicators that can support safe deployment. A video of the experiments is available at https://youtu.be/dN1Oes8pNIo.
Figures & tables
Fig. 1: Phases of the J-turn and the steering, pedal, and gear commands.
Fig. 2: Schematic of the vehicle’s planar motion as a symmetric bicycle model. df and dr are the clearances to the vehicles ahead and behind.
Failure Mode
Physical Cause
Evaluation Metric
Unsafe Control Action
Coverage Parameter
FM-1
Incomplete or excessive heading reversal
Heading change during maneuver
Premature or late phase transition
Heading threshold, reverse-speed threshold, initial state
FM-2
Tire force saturation
Peak lateral acceleration ay
Steering demand exceeds friction limit
Weather and road friction
FM-3
Yaw instability
Peak yaw rate r
Rotation grows beyond controllable region
Reverse-speed threshold, steering input
FM-4
Excessive roll
Peak roll angle ϕ
Load transfer reduces stability margin
Reverse-speed threshold, road friction
TABLE I: Physics-Based Failure Modes, Evaluation Metrics, and Coverage Dimensions
Fig. 3: (a) Outcome of the 1,152 trials. (b) Heading change during the maneuver; the shaded band is the 180∘±10∘ completion criterion.
Fig. 4: (a) Acceptance rate for combination of the phase-switching thresholds (96 trials/cell). (b) Acceptance rate by weather (with road friction μ ) and by pre-maneuver wait; the dashed line is the overall rate of 16.9%.
Fig. 5: Stability assessment of the J-turn ensemble. (a) Open-loop stability pyramid in (β,ψ˙,v) . (b) Classical phase portrait at vG=6 m/s. (c) The same data normalized by the bound at each sample’s own speed, so the admissible region is the unit square. Real traces use the same 3 s / 1.5 s / 1 s window as Fig. 7 b and are drawn at the ensemble speed so the overlay is geometric. Sideslip is wrapped into (−π/2,π/2) with the simulated body-frame convention. Shading is pooled sample density, the band is the 10th to 90th percentile, and the solid line is the ensemble median.
Real-1
Real-2
Sim. accepted
Ratio (Real/Sim)
Heading change
179.3∘
170.7∘
181.6∘
0.99 , 0.94
Peak ∣ay∣ (m/s 2 )
1.55
4.88
17.03
0.091 , 0.29
Peak yaw rate (°/s)
40.0
127.8
150.0
0.27 , 0.85
Peak roll (°)
2.63
1.94
0.246
10.7 , 7.9
TABLE II: Stunt-Driving Indicator Transfer Between Simulation and Real Vehicle
Fig. 6: G-Mini autonomous vehicle test-bed: roof-mounted perception unit, brake and throttle actuators, and modular steering robot.
Fig. 7: (a) Real-vehicle J-turn from aerial video, strobed at uniform intervals with the tracked path overlaid; the body rotates through the reversal while the path continues along its original direction, which produces the large sideslip. (b) Sideslip and yaw rate against maneuver time. Real traces share the simulated origin: 3 s of straight motion (cropped), the heading reversal compressed to 1.5 s, and 1 s of settled exit. Sideslip is folded by arcsin(sin(Δψ)) into (−π/2,π/2) ; yaw rate is converted from rad/s and scaled so its peak sits above 200∘ /s.
Recent advances in learning-based control have enabled impressive achievements in solving complex control problems in various domains. However, since learning-based control may not be able to realize safety-guaranties, it is of great importance to enhance safety and robustness while maintaining good performances. Take vehicle motion & dynamics control as an example, in order to overcome the pain points of traditional methods such as heavy parameter calibration effort and learning-based control to bring better performance and efficiency in stability & agility over prior work for state-based vehicle control tasks, in this work, our method aims to develop a curriculum learning controller enhanced with physics-based predictive safety filter. The validation is conducted with the Python-CarSim platform, demonstrating better improvements and scalability under various maneuvers.
Baocong Zhang, Siliang Lu, Chenyang Li
Shanghai Jiaotong University, Shanghai, China · Bosch Corporate Research, Shanghai, China
Training controllers that are safe and robust in simulation, and systematically assessing their readiness for real-world deployment, remain key challenges in sim-to-real transfer. To address this, we propose LyEvO, a physics-grounded framework that combines constrained Evolutionary Optimization and Statistical Model Checking (SMC)-based verification with Lyapunov-based stability analysis. Leveraging prior knowledge of the system dynamics, LyEvO uses Lyapunov analysis to compute an initial candidate stability region. An iterative loop then uses operational scenarios drawn from this region to jointly optimize and statistically verify a policy, and subsequently expands the region's boundaries based on the verification outcome. This integrated procedure provides a practical criterion for assessing deployment readiness. We evaluate LyEvO on Cartpole and 3D Quadrotor benchmarks through extensive simulations and targeted real-world experiments, demonstrating safe and robust sim-to-real transfer.
Riccardo Curcio, Hongpeng Cao, Marco Caccamo
Department of Computer Science, Sapienza University of Rome, via Salaria 113, 00198 Rome, Italy · School of Engineering and Design, Technical University of Munich, Boltzmannstraße 15, 85748 Garching b. München, Germany
Data-driven driving simulators command accelerations and steering rates from a fixed grid without constraining the realized accelerations and jerks. As a result, reinforcement-learning policies inflate safety metrics through abrupt, last-second maneuvers that lie far outside the range of human driving and would be unacceptable to occupants of a real vehicle, so the metrics measure simulator permissiveness rather than policy quality. Enforcing comfort bounds naively is not enough: lateral limits shrink quadratically with speed, so clamping a static grid saturates it and destroys fine-grained control ("grid collapse"). We propose an adaptive action parameterization that rediscretizes the grid at every step to span exactly the per-step feasible control set, via closed-form inversion of the lateral-jerk constraint. We further present PufferDrive-Editor, a browser-based tool to audit realized kinematics and author kinematically challenging scenes. On the Waymo Open Motion Dataset and a hand-authored slalom, our adaptive model holds comfort violations below 1% while outperforming clipped-grid and direct-jerk baselines in navigability.
Anna Rothenhäusler, Daniel Jost, Raghu Rajan +4
University of Freiburg, Freiburg, Germany · Bosch Center for Artificial Intelligence, Stuttgart, Germany · Coburg University of Applied Sciences, Coburg, Germany