△ Mechanism confirmed, baseline not beaten 2026

Polar-Backstepping Policy Residual

Implementation & benchmark of arXiv:2607.26442 — Global Exponential Stabilization of the Kinematic Bicycle Model of a Car in Polar Coordinates

Usefulness7/10

Difficulty5/10

Novelty7/10

Dynamics Stability Accuracy Rl World model Training dynamics Regularization Control theory Dynamical systems Geometry Differential geometry

Source paper: Global Exponential Stabilization of the Kinematic Bicycle Model of a Car in Polar Coordinates arXiv:2607.26442 ⓘ · analyzed Aug 31, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Represent car-like navigation states in the paper's polar coordinates and make a neural policy predict only a residual around an analytic backstepping controller. Add a Lyapunov-decrease penalty so the learned residual can improve trajectory quality without destroying the nominal parking attractor.

Formulas

$$\rho=\sqrt{x^2+y^2},\qquad \delta=\operatorname{atan2}(y,x)+\pi,\qquad \gamma=\delta-\theta$$

$$\dot\rho=-v\cos\gamma,\qquad \dot\gamma=\frac{v\sin\gamma}{\rho}-\frac{v}{L}\tan\phi,\qquad \dot v=a$$

$$V=\frac12\left(q_\rho\rho^2+q_\gamma\gamma^2+q_v(v-k_\rho\rho)^2\right),\qquad \mathcal L_{\mathrm{stab}}=\left[\dot V+\lambda V\right]_+^2$$

$$a=a_0+\Delta a_\vartheta(s),\qquad \phi=\operatorname{atan}\!\left(\tan\phi_0+\Delta u_{\vartheta}(s)\right)$$

Mathematical statement

For a target at the origin, define $\rho=\sqrt{x^2+y^2}$, $\delta=\operatorname{atan2}(y,x)+\pi$, and $\gamma=\delta-\theta$, where $\rho$ is range, $\delta$ is the line-of-sight angle, and $\gamma$ is heading error. Away from the coordinate cut and target singularity, the kinematic bicycle equations imply $\dot\rho=-v\cos\gamma$, $\dot\delta=v\sin\gamma/\rho$, $\dot\gamma=v\sin\gamma/\rho-(v/L)\tan\phi$, and $\dot v=a$, with speed $v$, wheelbase $L$, acceleration $a$, and steering angle $\phi$. Use a positive definite local Lyapunov function $V=\tfrac12(q_\rho\rho^2+q_\gamma\gamma^2+q_v(v-v_d(\rho))^2)$. Penalize positive drift $[\dot V+\lambda V]_+$, where $q_\rho,q_\gamma,q_v,\lambda>0$, so the desired certificate is $\dot V\le-\lambda V$. A bounded nominal choice is $v_d=k_\rho\rho$, $a_0=-k_v(v-v_d)-k_\rho v\cos\gamma$, and $\tan\phi_0=L\sin\gamma/(\rho+\epsilon)+Lk_\gamma\gamma/(v+\epsilon)$, with positive gains $k_\rho,k_v,k_\gamma$ and safeguard $\epsilon>0$.

Implementation notes

1. Integration point: use this controller as a reinforcement-learning or imitation-learning policy for a kinematic bicycle simulator. At every step compute $s=(\rho,\sin\gamma,\cos\gamma,v,1/\rho_\epsilon)$, where $\rho_\epsilon=\max(\rho,\epsilon)$. The network outputs bounded residuals $(\Delta a,\Delta u)$ through tanh heads. The analytic controller computes $(a_0,\phi_0)$, and the final actions are clipped to actuator limits. 2. Pseudocode: compute the polar state; compute $v_d=k_\rho\rho$; compute $a_0=-k_v(v-v_d)-k_\rho v\cos\gamma$; compute $u_{\phi0}=L\sin\gamma/\rho_\epsilon+Lk_\gamma\gamma/(|v|+\epsilon)$; evaluate the MLP residual; apply the actions; estimate $\dot\rho,\dot\gamma,\dot v$ from the analytic equations or finite differences; calculate $\dot V$; and optimize task loss plus $\beta[\dot V+\lambda V]_+^2$. 3. Computed from the paper: the polar geometry, transformed dynamics, and strict-feedback structure. Estimated empirically: $\lambda$, gains, residual bounds, and the fraction of sampled states satisfying the certificate. 4. First cheap experiment: train a two-layer MLP on randomly generated parking goals in a bicycle simulator, comparing residual training against a Cartesian policy and an unconstrained polar policy. Use $L=2.5$, $\epsilon=0.05$, and measure $V(t)$, success rate, and the largest initial-error radius for which $V(t+1)\le e^{-\lambda\Delta t}V(t)$ holds. The quantitative prediction is an approximately linear semilog decay of $V$ with slope no greater than $-\lambda$ over the certified region. Removing the stability loss should sharply increase positive-$\dot V$ events, particularly when $|v|<0.2$ or $\rho<0.5$.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Significantly worse ✗

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built a fixed-seed polar backstepping residual-policy MVP with bounded MLP residuals, Lyapunov drift penalty, exact polar dynamics, and three parameter sweeps. All three mechanism predictions were confirmed: speed-error decay matched k_v with identity error below 2.3e-16, steering decay matched k_gamma*v/(|v|+epsilon) with relative error below 6.2e-15, and the epsilon transition matched the predicted 1/r-1/epsilon residual below epsilon while vanishing above it. In the toy policy comparison, the Lyapunov penalty reduced positive-drift states from 13.09% to 10.89% and improved mean drift-plus-lambda-V from -0.135 to -0.206, but increased task MSE from 0.00071 to 0.0611, so the stability effect is real but comes with a clear accuracy tradeoff.

Agent confidence: 8/10

Baseline: Unconstrained polar residual: task MSE 0.0007122, positive drift fraction 0.13086, mean (dV+lambda V) -0.13484
Idea: Lyapunov-regularized polar residual: task MSE 0.06113, positive drift fraction 0.10889, mean (dV+lambda V) -0.20551; 16.8% relative reduction in positive-drift events

Limitations: This is a sampled one-step toy study rather than full bicycle trajectory navigation; no success-rate or long-horizon parking evaluation was performed, and no Cartesian-policy baseline was implemented. The certificate is evaluated on sampled states and uses clipped actions, so it is not a formal global stability guarantee.

How to run: python3 polar_backstepping_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.26442) Sep 2, 2026

Significantly WORSE than the tuned baseline (+917.7%, p=0.0081, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0004
Idea mean: 0.004
Effect (Δ): 0.0036 (+917.7%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±334.1%
Mechanism: Confirmed ✓
Practical effect: Hurts
Baseline tuning: swept over 6 configs

Limitations:

The built-in dynamics track predicts a future pendulum angle rather than directly learning a car-like navigation action policy, so the intervention uses a one-step finite-difference pendulum Lyapunov proxy adapted to the available control structure. No long-horizon closed-loop parking, actuator clipping study, Cartesian policy comparison, or larger-scale neural architecture was tested.

How to run: python3 stage2_bench.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Aug 31, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented the Polar-Backstepping Lyapunov residual as a modified training loss on the structurally matched dynamics track using the shared rnn_small architecture. Across 8 paired seeds, the idea significantly worsened test MSE (delta +0.0036311, permutation p=0.0081) despite confirming its mechanism signature: positive drift fraction fell by 22.3% (0.611 to 0.475). The benchmark therefore rejects the idea as a task-performance win.

Baseline: Best sweep lr=0.01, weight_decay=0.0; 8-seed test MSE mean 0.000395668 ± 0.000148467
Idea: Best beta=0.1 of {0.1,0.5,1.0}, lr=0.01, weight_decay=0.0; 8-seed test MSE mean 0.00402676 ± 0.00157548; paired delta +0.00363109; p=0.0081; positive-drift fraction 0.4747 vs baseline 0.6109

Benchmark result