A car whose four wheels are each driven by their own electric motor has more controls than it strictly needs. Four wheel torques decide only three things about how the body moves: the forward push, the sideways force, and the turning moment. That spare degree of freedom is usually spent on efficiency or ignored. We propose to spend it on learning safely. The vehicle can deliberately push one tire toward its grip limit while the other three wheels cancel the effect on the body. This reveals how much friction the road offers, and it does so without the chassis ever leaving its safe envelope. The friction estimate then widens a safety filter that lets a reinforcement-learning policy explore closer to the true handling limit, where drifting and emergency maneuvers live.
This matters because current safe-exploration methods face a known trap at the handling limit. Once a learner becomes safe, it stops seeing what unsafe looks like, so its estimate of where the boundary lies decays. The fix in Breaking Safety Paradox with Feasible Dual Policy Iteration is to run a second, deliberately reckless policy that violates constraints, which no one can afford on a physical vehicle near spin-out. Over-actuated vehicles offer a different exit. They can collect boundary information at the level of a single actuator and tire while the body stays inside its constraints.
Core Claim
In a vehicle with independent wheel torque control, excitations confined to the approximate null space of the torque-to-body-force map can identify per-wheel friction limits. Each excitation is designed so the predicted change in body motion stays below a set bound. A safety filter that uses the resulting friction confidence bounds can then be relaxed online, giving faster recovery of near-limit performance after surface changes, with fewer chassis-level violations than either robust fixed filters or random excitation. The mechanism should fail in a specific, testable way on rear-wheel drive, which has no redundancy, and that failure is our main falsification check.
Concretely, let $u \in \mathbb{R}^4$ be the wheel torques and $w = (F_x, F_y, M_z)$ the resulting body wrench (forward force, lateral force, yaw moment). Write $J(s)$ for the $3\times 4$ Jacobian of $w$ with respect to $u$ at the current state $s$, computed from the current tire model. A probe is a perturbation
$$u = u_\pi + n(s)\,z, \qquad J(s)\,n(s) = 0,$$
where $u_\pi$ is the policy's command, $n(s)$ spans the one-dimensional null space, and $z$ is a scalar probe amplitude. Tires are nonlinear and the friction circle couples longitudinal and lateral force, so the null space is only locally exact. We therefore cap $z$ so that a second-order prediction of the wrench change $\|\Delta w\|$ stays under a bound $\delta$. The closed-loop allocator absorbs the remainder. The probe drives one wheel's slip ratio toward the peak of its force curve. The measured wheel acceleration against commanded torque then gives a direct reading of that wheel's peak friction coefficient $\mu_i$.
Research Program
The program has three parts. The first is an independent-wheel-drive simulator that runs in parallel on a GPU, following the setup of Learning to Drift with Individual Wheel Drive. That paper trained a PPO drifting policy on a 1/10-scale four-motor car in about 10.7 minutes and transferred it zero-shot to the real vehicle. We extend the simulator with the effects that paper explicitly omits: lateral load transfer between inner and outer wheels, motor torque-rate limits, and spatially varying friction patches, such as a dry floor giving way to wet or icy sections mid-maneuver. Load transfer matters here because it changes which wheel can be probed cheaply.
The second part is the probing module and a friction posterior. The posterior holds a per-wheel Bayesian estimate of $\mu_i$, updated from probe responses and from ordinary driving data. A scheduler decides when to probe. It probes when posterior uncertainty about the wheel that limits the current maneuver is high, and when the predicted wrench disturbance is small.
The third part is a predictive safety filter. Over a short horizon, it projects the policy's action onto the set whose rollouts keep track error, sideslip rate and yaw rate inside the envelope, under the lower confidence bound of the friction estimates. Because tire forces enter non-affinely, the vertex trick that handles input limits in Safe Control Under Input Limits with Neural Control Barrier Functions does not apply directly. That paper learns barrier certificates for a 10-dimensional quadcopter-pendulum with polytopic, control-affine inputs. We borrow its counterexample-search idea to validate our filter offline, but implement the filter itself as a sampling-based short-horizon check.
Inside the filter, the policy is trained with a boundary-seeking update, following Boundary-Seeking Policy Gradient for Safe Reinforcement Learning. That method splits the policy update into a part along the constraint boundary and a part pulling toward it, and it argues the optimal constrained policy sits on the boundary. We use it because drift stabilization is a boundary-hugging task. Our contribution is to make that boundary safe to approach by shrinking the friction uncertainty that otherwise forces the filter to be conservative.
The tension with conservative state-wise bounds, as in Constrained Markov Decision Processes via Backward Value Functions, is addressed directly. Conservatism there comes from upper-bounding trajectory cost. Here it comes from friction uncertainty, which probing reduces. We expect, and will measure, the filter's admissible envelope growing as posterior variance falls.
Baselines
The do-nothing reference is the Zhou et al. policy: PPO trained with domain randomization over tire parameters and deployed with no online adaptation and no filter. This is the strongest existing result on this platform. On hardware it reports a position RMSE of 0.22 m on an eight-shaped trajectory and a mean sideslip of about 47°, according to Learning to Drift with Individual Wheel Drive. It is the right competitor because its authors argue that broad randomization makes online safety machinery unnecessary. We retrain it in our extended simulator with the same randomization ranges, extended to cover our friction patches, and the same GPU budget. Beating it means fewer spin-outs and lower tracking error after unannounced friction transitions, at equal or better nominal-surface tracking.
The strongest safe-exploration competitor is constrained online fine-tuning with PPO-Lagrangian, starting from the same pretrained policy. Lagrangian PPO is the reference constrained method on Safety Gymnasium, which reports a 98% reduction in velocity-constraint violations against plain PPO at a 45% reward cost. We also include a variant pretrained with FDPI in simulation and then fine-tuned conservatively. FDPI reports near-zero violations across 14 Safety-Gymnasium tasks, according to Breaking Safety Paradox with Feasible Dual Policy Iteration, but it explores by violating constraints, so it cannot probe online on hardware. On safe exploration for independent-wheel torque vectoring specifically, there is no reported result on Lacuna, so our numbers there are first measurements rather than comparisons to published figures.
The robust reference is our own filter run with a fixed worst-case friction value and no learning of $\mu$. Comparing against it isolates whether learning friction helps at all.
The key ablation replaces null-space probing with isotropic random torque dithering of matched energy, using the same estimator and filter. It isolates whether redundancy, rather than mere excitation, is what makes probing safe. A second ablation applies the full method to a rear-wheel-drive configuration of the same car. That configuration has a two-input drive, and with steering taken as fixed the drive torques leave no null space, so probing necessarily disturbs the body.
Every method gets the same simulator seeds, the same friction-transition schedules, the same number of online interaction steps, and the same hyperparameter search budget of 20 configurations each.
Why not just…?
Why not just randomize friction harder in simulation and deploy zero-shot? This is the most serious threat, since Learning to Drift with Individual Wheel Drive shows strong transfer and reports that removing initial-state randomization cost 29% in success rate. Randomization matters a great deal. However, a policy randomized over all friction values must hedge, behaving reasonably everywhere and optimally nowhere. It also has no way to notice that this particular surface is icy. Our friction-transition experiment settles the question directly. If the randomized policy matches our spin-out rate and tracking after transitions, the probing story is unnecessary and we will report that.
Why not just estimate friction from ordinary driving, or from random dithering? During steady cornering, tires rarely sit near their peak, so passive estimates are poorly identified exactly where the limit matters. Random dithering does excite the tire, but it also shakes the body. The matched-energy dithering ablation measures both the wrench disturbance and the estimation error. If dithering gives comparable identification with comparable chassis disturbance, then the null-space projection adds nothing, and the rear-wheel-drive ablation will show the same pattern.
Why not just use MPC with an adaptive tire model? Model predictive control is the natural classical answer, and our filter is a short-horizon predictive check. But a predictive controller still needs informative data about the limit, and it gets that data only by approaching the limit. The physics-informed diffusion model in One Model to Drift Them All feeds learned dynamics parameters to a 200 Hz MPC and reports 1.47 m lateral error in heavy rain. It learns those parameters offline from 84 pre-collected runs, which presumes someone already drove at the limit. Probing is complementary to MPC. We will also report an adaptive MPC variant that uses our probe-fed friction posterior, to separate the value of the data from the value of the learned policy.
Why not just a static torque or sideslip cap? A cap below about 47° of sideslip forbids the drift itself. Neurosymbolic Reinforcement Learning with Formally Verified Exploration found static shields too restrictive, outperformed by its evolving shield on 4 of 10 tasks. The fixed-friction robust baseline tests this directly.
Proposed Experiments
The main simulation study runs eight-shaped and circular drift-tracking tasks, with friction patches introduced at random times and locations and with load transfer enabled. We measure chassis constraint violations, separating spin-outs from track departures, cumulative over online learning. We also measure tracking RMSE and mean sideslip after each transition, the time for per-wheel friction error to fall below 10%, and the body-wrench disturbance caused by probing.
The first decisive check compares four-wheel independent drive against rear-wheel drive. If probing helps equally without redundancy, our explanation is wrong.
The second decisive check plots the filter's admissible sideslip envelope against posterior friction variance. The claimed mechanism predicts that the envelope widens monotonically as variance falls, and that tracking error and violations improve once it does.
The third check tests robustness to the locally linear null space. We sweep the disturbance bound $\delta$ and verify that realized body disturbance tracks it at high slip, where the Jacobian approximation is weakest.
If simulation results hold, we deploy on a 1/10-scale four-motor car on two floor surfaces, one of them lined with a low-friction sheet. We compare spin-out counts during a fixed budget of online adaptation laps against the zero-shot randomized policy.
References
- Breaking Safety Paradox with Feasible Dual Policy Iteration
- Learning to Drift With Individual Wheel Drive: Maneuvering Autonomous Vehicle at the Handling Limits
- Safe Control Under Input Limits with Neural Control Barrier Functions
- Boundary-Seeking Policy Gradient for Safe Reinforcement Learning
- Constrained Markov Decision Processes via Backward Value Functions
- Safety Gymnasium: A Unified Safe Reinforcement Learning Benchmark
- One Model to Drift Them All: Physics-Informed Conditional Diffusion Model for Driving at the Limits
- Neurosymbolic Reinforcement Learning with Formally Verified Exploration