✗ Failed on benchmark 2026

Reachable-Set Risk Head for Early-Warning Rollouts

Implementation & benchmark of arXiv:2607.26527 — Online Monitoring and Risk Assessment of Non-Cooperative UAVs via STL-Aware Adaptive Fusion Kalman Filtering

Usefulness7/10

Difficulty6/10

Novelty5/10

Dynamics Accuracy Stability Generalization World model Rnn Ssm Training dynamics Regularization Dynamical systems Control theory Probability Statistics

Source paper: Online Monitoring and Risk Assessment of Non-Cooperative UAVs via STL-Aware Adaptive Fusion Kalman Filtering arXiv:2607.26527 ⓘ · analyzed Aug 31, 2026

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 Not peer-reviewed.

Idea description

Attach a probabilistic reachable-set head to a neural world model so that long-horizon predictions produce both a mean trajectory and an uncertainty envelope. Train or calibrate the model using the probability that the predicted envelope intersects an unsafe region, allowing early-warning losses to penalize risk before an actual violation appears. The transferable signature is a predictable monotone increase in warning probability as the reachable set approaches or intersects a forbidden set.

Formulas

$$\mu_{t+h+1}=F_\theta(\mu_{t+h},u_{t+h}),\qquad \Sigma_{t+h+1}=J_{t+h}\Sigma_{t+h}J_{t+h}^{\mathsf T}+Q_\theta,$$

$$q_{r,h}=1-\Phi\!\left(\frac{b_r-c_r^{\mathsf T}\mu_{t+h}}{\sqrt{c_r^{\mathsf T}\Sigma_{t+h}c_r}}\right),\qquad R_t=1-\prod_{h=1}^{H}\prod_r(1-q_{r,h}),$$

$$\mathcal{L}_{\mathrm{total}}=\mathcal{L}_{\mathrm{forecast}}+\lambda_R\operatorname{BCE}(R_t,v_t)+\lambda_C\left|\operatorname{PCE}(R_t)-\alpha\right|,$$

Mathematical statement

Let the learned latent dynamics be x_{t+1}=F_theta(x_t,u_t)+epsilon_t, with Gaussian approximation x_{t+h} distributed as N(mu_{t+h},Sigma_{t+h}). Linearize F_theta around the predicted mean with Jacobian J_{t+h}=partial F_theta/partial x, and propagate covariance by Sigma_{t+h+1}=J_{t+h}Sigma_{t+h}J_{t+h}^T+Q_theta. For an unsafe half-space c_r^T x >= b_r, its violation probability is q_{r,h}=1-Phi((b_r-c_r^T mu_{t+h})/sqrt(c_r^T Sigma_{t+h} c_r)), where Phi is the standard normal CDF. A finite-horizon risk score is R_t=1-product over h and r of (1-q_{r,h}). Train with binary cross-entropy between R_t and the indicator of a future violation, together with calibration loss. The mechanism is uncertainty growth through Jacobian products and an explicit probabilistic intersection test.

Implementation notes

Add this head to a compact GRU, SSM, or Transformer world model used for multi-step forecasting. The model should output a state mean and either diagonal process covariance Q_theta or a low-rank-plus-diagonal covariance. Define unsafe regions as half-spaces, or approximate nonlinear regions by multiple local half-spaces. Initialize mu=x_t and Sigma from the encoder; for each forecast step, compute mu=F_theta(mu,u), obtain the Jacobian using autograd or Jacobian-vector products, propagate Sigma, evaluate every q_{r,h}, and accumulate R=1-product(1-q). Add binary cross-entropy against whether the ground-truth future enters any unsafe set. Initially stop gradients through the Jacobian if memory is excessive. The covariance recursion and Gaussian tail probability come from the mechanism; Q_theta, initial covariance, covariance inflation, and calibration temperature are estimated from held-out residuals. The first cheap experiment should use a 2-D synthetic constant-turning and acceleration dataset with noisy, partially observed trajectories and a GRU trained for 20-step prediction. Compare deterministic prediction, deep ensembles, and this reachable-set head using violation recall at fixed 5% false-warning rate and expected calibration error. The prediction is that R_t increases monotonically as the Mahalanobis safety margin m=(b-c^T mu)/sqrt(c^T Sigma c) decreases, with a transition near m=0; empirical violation frequencies binned by predicted R_t should track R_t, and uncertainty growth should follow the dominant Jacobian-product singular value.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

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 and numerically verified the reachable-set risk head for Gaussian linearized dynamics. The predicted monotone margin response, q=0.5 transition at zero margin, stable-versus-unstable covariance behavior, and increasing horizon union risk all matched quantitatively. In the 2-D rollout experiment, reachable risk had recall 0.710 versus 0.733 for a fair endpoint baseline at 5% false-warning rate, but its ECE was 0.117 versus 0.081; therefore the mechanism manifested, but no overall warning-quality win was observed.

Agent confidence: 8/10

Baseline: Fixed-variance endpoint Gaussian warning: recall@5% FPR=0.733, ECE=0.0812
Idea: Reachable-set Gaussian union risk: recall@5% FPR=0.710, ECE=0.1169; mechanism checks: q(m=0)=0.5 observed, stable covariance limit error=3.9e-14, horizon-5 union risk=0.40951 observed

Limitations: No learned GRU, calibration loss, deep ensemble, nonlinear dynamics, multiple half-spaces, or held-out calibration procedure was tested. The toy dynamics were linear and the reachable score used known covariance and independent half-space/horizon risk multiplication.

How to run: python3 run_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.26527) Sep 2, 2026

Significantly WORSE than the tuned baseline (+67763.2%, p=0.0081, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0021
Idea mean: 1.4291
Effect (Δ): 1.427 (+67763.2%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±4266.2%
Mechanism: Not confirmed ✗
Practical effect: Hurts
Baseline tuning: swept over 3 configs

Limitations:

The built-in dynamics benchmark is one-step pendulum regression, so the implementation used a next-state safety margin and finite-horizon union transform rather than full multi-step autograd Jacobian covariance propagation. Calibration fitting, temperature scaling, ensembles, multiple half-spaces, and long-horizon warning recall were not tested. The custom idea path fell back to CPU after a shared-GPU allocation failure.

How to run: python3 bench_reachable.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: 10/10

The reachable-set risk head was implemented on the registered dynamics/rnn_small benchmark with a matched GRU backbone and an added Gaussian risk-training loss. It significantly worsened standard test MSE relative to the tuned baseline: paired delta +1.42697 with permutation p=0.0081, so the benchmark verdict is idea worse (significant). The trained-model mechanism signature was also not confirmed quantitatively.

Baseline: dynamics/rnn_small, best lr=0.006, full 8-seed test MSE mean 0.002105819425196387
Idea: reachable-risk GRU, best lr=0.001 and risk_weight=0.08, full 8-seed test MSE mean 1.4290764033794403

Benchmark result