# Rigidity-Conditioned Active-Sensing Policy

- ID: 717
- Canonical URL: https://synthcore.org/idea/717/rigidity-conditioned-active-sensing-policy
- API JSON: https://synthcore.org/api/idea/717.json
- API Markdown: https://synthcore.org/api/idea/717.md
- Verification status: unverified
- Source: [arXiv:2607.10933](https://arxiv.org/abs/2607.10933)
- Category: geometry
- Solves: stability, sample-efficiency, accuracy
- ML areas: rl, world-model, training-dynamics
- Math tags: linear-algebra, geometry, graph-theory, control-theory, optimization
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Add a differentiable geometric-conditioning reward to a neural policy that selects UAV motions or other active-sensing actions. The policy is rewarded for configurations whose sensing Jacobian has a large smallest nonzero singular value, preventing early decisions from overfitting to an uncertain target estimate and encouraging measurements that distinguish competing hypotheses.

## Mathematical statement

Let $p_i\in\mathbb{R}^d$ be the position of sensor or UAV $i$, let $q\in\mathbb{R}^d$ be the target position, and let $G=(V,E)$ be the sensing graph. For an edge $e=(i,j)$, define the relative direction $u_{ij}=(p_i-p_j)/\|p_i-p_j\|$. The rigidity matrix $R(p)$ has one row per edge and $d$ columns per node, with block $u_{ij}^{\mathsf T}$ in the columns of node $i$, block $-u_{ij}^{\mathsf T}$ in the columns of node $j$, and zeros elsewhere. For a target-sensing edge $(i,q)$, the same construction uses $p_q=q$. Let the singular values of $R$ be $\sigma_1\geq\cdots\geq\sigma_n\geq0$ and let $r$ be the expected rank after removing gauge freedoms such as global translation and rotation. The rigidity score is $\rho(R)=\sigma_r(R)$, the smallest nonzero singular value. The policy maximizes expected cumulative reward augmented by $\beta\log(\rho(R)+\epsilon)$, where $\beta>0$ and $\epsilon>0$ prevent numerical singularities. The mechanism predicts poor local observability when $\rho$ approaches zero and improved robustness when $\rho$ increases.

## Key formulas

- $$R_{e,:}=\begin{cases}u_{ij}^{\mathsf T}\text{ in the block of node }i,\\-u_{ij}^{\mathsf T}\text{ in the block of node }j,\\0\text{ otherwise,}\end{cases}\qquad u_{ij}=\frac{p_i-p_j}{\|p_i-p_j\|_2}.$$
- $$\rho(R)=\sigma_r(R),\qquad \mathcal{J}(\theta)=\mathbb{E}_{\pi_\theta}\left[\sum_{t=0}^{T-1}\gamma^t\left(r_t+\beta\log(\rho(R_t)+\epsilon)\right)\right].$$
- $$\widetilde{R}_t=R_t[\mathcal{E}_t],\qquad \rho_k(R_t)=\sigma_r(\widetilde{R}_t)$$

## Implementation notes

Integrate this at the reward and auxiliary-loss interface of an actor-critic or PPO policy. The actor receives UAV positions, noisy target hypotheses or a particle set of target hypotheses, and recent measurements; it outputs headings or velocity commands. A differentiable simulator predicts the next positions and the corresponding range, bearing, or time-difference-of-arrival sensing graph. Construct one rigidity-matrix row for every selected sensor-target or sensor-sensor relation, compute the singular values with `torch.linalg.svdvals`, and add `beta * log(sigma_r + epsilon)` to the reward. If target uncertainty is represented by particles, compute the score for each particle and use the minimum or a soft minimum, `-tau*logsumexp(-rho/tau)`, so the policy cannot exploit only one optimistic hypothesis. To reduce cost, use a fixed-size edge subset `E_t` containing longest-baseline, nearest, and highest-quality links; this is the pruning idea, while the full matrix is retained for evaluation. The paper-derived quantities are the rigidity construction and smallest-nonzero-singular-value criterion. Estimate empirically the gauge rank `r`, sensor noise, and reward coefficient `beta`; for a known 2-D localization setup, explicitly project out known translation and rotation nullspaces or calibrate `r` from noiseless configurations. First run a cheap experiment with 4--8 simulated UAVs, a moving target, range/bearing noise, and random initial target error. Compare PPO with task reward alone, PPO plus determinant/FIM reward, and PPO plus rigidity reward. Measure localization error, time to enter the 50 m radius, and the singular spectrum. The quantitative prediction is a sharp degradation boundary: episodes with `rho(R_t)` near zero should have substantially larger localization error, while conditioning improves monotonically as `rho` rises. With pruning, the rank and score should remain within 10--20% of the full matrix while SVD computation decreases roughly in proportion to the number of retained edges. Sweep policy actions to estimate the minimum score threshold `rho_c` at which the 80% success requirement is crossed; the rigidity method predicts this transition should be more stable under target-estimate bias than an FIM-only policy.

## Disclaimer

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