△ Mechanism confirmed, baseline not beaten 2026

Differentiable Separating-Axis Clearance Barrier

Implementation & benchmark of arXiv:2608.26533 — Barrier Function Conformal Safety Clearance Certification with CVaR for Driving Trajectory Selection

Usefulness8/10

Difficulty5/10

Novelty6/10

Regularization Stability Accuracy Generalization Optimizer Rl World model Geometry Optimization Control theory

Source paper: Barrier Function Conformal Safety Clearance Certification with CVaR for Driving Trajectory Selection arXiv:2608.26533 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Add a geometric barrier loss to a neural trajectory generator or scorer using separating-axis margins between ego and predicted-agent oriented bounding boxes. The barrier is differentiable almost everywhere and has direct collision meaning, unlike an arbitrary proximity penalty.

Formulas

$$g_n=|n^T(c_B-c_A)|-r_A(n)-r_B(n),\qquad b(A,B)=\max_{n\in\{u_A,v_A,u_B,v_B\}}g_n.$$

$$\widetilde b_\tau=\tau\log\left(\sum_n\exp(g_n/\tau)\right),\qquad \mathcal{L}_{barrier}=\sum_{t,j}\operatorname{softplus}\left((m_0-\widetilde b_{\tau,tj})/\beta\right).$$

Mathematical statement

For rectangles A and B, let c_A,c_B be centers, u_A,v_A,u_B,v_B unit body axes, and a_A,b_A,a_B,b_B half-lengths and half-widths. For each separating-axis candidate n in {u_A,v_A,u_B,v_B}, the projected half-width is r_A(n)=a_A|u_A^T n|+b_A|v_A^T n| and similarly for B. The signed projection gap is g_n=|n^T(c_B-c_A)|-r_A(n)-r_B(n). The separating-axis margin is b(A,B)=max_n g_n. For separated convex rectangles, positive b is a conservative lower bound on Euclidean clearance; b<=0 indicates that no tested axis certifies separation. Smooth absolute values and maxima permit backpropagation.

Implementation notes

1. Integration point: make a neural planner output ego poses and box dimensions over a horizon, or output a score for each candidate trajectory. Obtain surrounding-agent boxes from logged future states, a learned predictor, or a simulator. Compute the barrier at every time and for every agent, and add lambda_b times the barrier loss to imitation learning, supervised trajectory prediction, or the RL objective. At inference, report the minimum barrier and reject or rerank candidates below threshold m_0. 2. Pseudocode: construct the four axes from the two box orientations; compute smoothed absolute projections, projected radii, and g_n; compute the log-sum-exp margin; accumulate softplus((m_0-margin)/beta); backpropagate through the trajectory network. Start with tau=0.1 meters and beta=0.05 meters. 3. Computed from the mechanism: axes, projections, radii, barrier values, and gradients. Estimated empirically: lambda_b, m_0, beta, and prediction error in other-agent boxes. 4. First experiment: train a small MLP or transformer on nuScenes or nuPlan with and without this loss, then evaluate exact polygon or OBB signed clearance independently. The predicted signature is b(A,B)<=d_exact(A,B) for separated boxes and a sharp increase in collision rate when the minimum barrier crosses zero. At least 95 percent of positive-barrier cases should have nonnegative exact clearance, and the smoothed zero crossing should be within 0.1 meters of the exact SAT crossing.

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 differentiable oriented-box SAT barrier with smooth absolute projections, log-sum-exp margin, and softplus clearance loss. The exact axis-aligned zero crossing occurred at 2.00 m as predicted, while the tau=0.1 smoothed crossing was 1.95 m (0.05 m bias, within the 0.1 m target); smoothing errors stayed below tau*log(4), and gradients approached 1.0 on the active separating axis. In the tiny planner experiment, the barrier reduced collision fraction from 88.9% to 3.5% and improved minimum margin from -0.998 to -0.133, but increased imitation MSE from 2.7e-5 to 0.309, so the safety effect is real but not a Pareto win under this weighting.

Agent confidence: 8/10

Baseline: MSE 2.71e-05; minimum SAT margin -0.998 m; collision fraction 88.9%
Idea: MSE 0.309; minimum SAT margin -0.133 m; collision fraction 3.52%; exact crossing 2.00 m; smoothed crossing 1.95 m; smoothing bound and gradients verified

Limitations: Only a synthetic axis-aligned obstacle and a small MLP were tested; no nuScenes/nuPlan data, rotated-box polygon benchmark, multi-agent forecasting, hyperparameter sweep, FLOP-normalized comparison, or inference reranking evaluation was performed.

How to run: python3 barrier_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.26533) Sep 2, 2026

Significantly WORSE than the tuned baseline (+67.9%, p=0.0242, wins 1 / 8 paired seeds).

Benchmark: Obb_clearance_trajectory
Model: mlp_tiny
Paired seeds: 8
Baseline mean: 0.1176
Idea mean: 0.1974
Effect (Δ): 0.0798 (+67.9%; negative = idea better)
Wins: 1 / 8 paired seeds
p-value: 0.0242 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: Hurts
Baseline tuning: swept over 3 configs

Limitations:

The custom track is synthetic rather than nuScenes/nuPlan and uses a small MLP and 24 epochs. Exact polygon clearance, multiple agents, inference reranking, FLOP-normalized comparisons, and broader barrier-weight sweeps were not tested.

How to run: python3 obb_bench.py

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

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 9/10

Built and ran a structurally matched custom OBB trajectory-regression track with identical MLP systems, paired 8-seed evaluation, equal learning-rate sweeps, differentiable SAT barrier training, and a trained-model clearance signature. The barrier reduced observed collision fraction to 0.3026 and improved predicted margins, but test MSE increased from 0.11758 to 0.19739; paired delta was +0.07982 with permutation p=0.02415. The safety mechanism is directionally present, but the idea is significantly worse on the standard task metric and is not a benchmark win.

Baseline: MSE 0.1175788217 mean, std 0.0498898437, best lr 0.01, 8 seeds
Idea: MSE 0.1973940274 mean, std 0.1311616229, best lr 0.01, lambda 0.08, tau 0.1, m0 0.05, beta 0.05; collision fraction 0.3025781

Benchmark result