Safety verification for drones and spacecraft asks a concrete question: from which orientations and spin rates can the vehicle recover, whatever the disturbances do? The standard tool is Hamilton-Jacobi reachability. It computes a "value function" whose sign at each state says whether that state is recoverable, by solving a partial differential equation on a grid. That grid grows exponentially with the number of state variables, which is the curse of dimensionality. A second problem compounds it. Every solver that copes with the dimension problem treats the state as a flat box of numbers, so orientation gets written as three Euler angles (roll, pitch, yaw). At 90° of pitch those angles hit gimbal lock: two rotation axes line up, one degree of freedom vanishes, and derivatives blow up. The usual fix is to forbid large tilts, which removes exactly the flips and tumble recoveries where a safety certificate matters most.
This note proposes a learn-then-repair pipeline that works on the space of rotations itself rather than on any flat chart of it. A neural network learns an approximate value function on rotations and angular velocities. Grid-based dynamic programming then corrects that network only in a thin band around its predicted safety boundary. The grid lives on the rotation group, and each update step moves along true rotations, so no coordinate singularity can appear. Two outputs make this concrete. The first is a reach-avoid set ("reach the target while avoiding failure") covering all orientations, certified relative to a grid dynamic program, for a 6-dimensional attitude system whose body axes are strongly coupled. The second is a measurement showing where gimbal-lock artifacts corrupt Euler-angle reachability solutions, and by how much.
Core Claim
The nearest prior work on attitude reachability is ConReach (Xu et al.). It bounds reachable sets on $SO(3)\times\mathbb{R}^3$ by geodesic "metric balls", using contraction theory and a semidefinite program. Its proofs, however, require the initial set to be small enough to be strongly convex on the manifold, and covering the whole rotation group is left to future work. Its output is a forward over-approximation, not a controlled reach-avoid set. The high-dimensional neural solvers, DeepReach (Bansal & Tomlin) and its parametric FaSTrack extension (Jeong et al.), scale to 9–13 dimensions but carry no worst-case guarantee and use flat orientation coordinates.
Our claim is that full-envelope attitude safety sets can be computed and certified at 6 dimensions if two things change. Certification must be done by local dynamic programming on an orientation grid that stays faithful to the rotation group. The expensive certified computation must also be confined to the boundary band of a learned guess. We expect the dominant errors of flat-coordinate solvers to concentrate near gimbal lock, and we expect the band-restricted repair to remove them at a small fraction of full-grid cost.
One point should be said up front, because it shapes what is and is not new here. Feeding the nine entries of the rotation matrix into a DeepReach-style network already gives a value function that is free of singularities and does not represent any orientation twice. The chain rule automatically pairs the network's gradient with valid rotational velocities only. That ingredient is simple, and we adopt it as our learner rather than claim it. The contribution is the certified repair on the rotation group, the full-envelope benchmarks with ground truth, and the measured comparison against chart-based alternatives.
Research Program
The state is $(R,\omega)$, where $R$ is a rotation matrix and $\omega$ is the body-frame angular velocity. The dynamics are $\dot R = R\hat\omega$ and $\dot\omega = J^{-1}(\tau + d - \omega\times J\omega)$. Here $\hat\omega$ is the skew-symmetric matrix built from $\omega$, $J$ is the inertia matrix, $\tau$ is a bounded control torque, and $d$ is a bounded disturbance torque. The cross-product term is the gyroscopic coupling, which is why decomposition into independent axes (Chen & Tomlin) does not apply away from hover.
The learner is a sinusoidal network, as in DeepReach, with inputs $(\mathrm{vec}(R),\omega,t)$. It is trained on the standard reachability variational-inequality loss, with training states sampled uniformly over rotations instead of over an angle box.
The repair stage follows HJ-PATCH (Tonkens & Herbert). That method fixes a learned barrier by running dynamic programming only on states that violate invariance, and it reported a 100× compute reduction on a 6D planar quadrotor. Its grid and neighbor updates, though, are Euclidean. We replace them with a near-uniform grid on $SO(3)$, built by layering circle samples over a sphere grid, which avoids the polar crowding of an Euler-angle grid. Each grid node is updated with a semi-Lagrangian step that moves along the group exponential:
$$V_{k+1}(R,\omega)=\min\Big\{\ell(R,\omega),\ \max_{\tau}\min_{d}\ \mathcal{I}V_k\big(R\exp(\Delta t\,\hat\omega),\ \omega+\Delta t\,J^{-1}(\tau+d-\omega\times J\omega)\big)\Big\}.$$
In this update, $\ell$ is the signed distance to the failure set, $\Delta t$ is the time step, and $\mathcal{I}$ interpolates the value function on the $SO(3)$ grid using geodesic-distance weights. Because each step is an exact rotation, the integrator does not drift off the set of valid rotations. This is the property that variational integrators on $SO(3)$ (Hussein et al.) maintain, and that projected quaternion integration loses.
Updates are restricted to cells within a band of the learned zero level set. The band is grown until no boundary cell changes sign, as in HJ-PATCH's local propagation. The resulting guarantee is deterministic relative to the grid dynamic program. As a separate statistical layer, we add the conformal-prediction safety margin of Tayal et al. over continuous rollouts.
When the failure set and dynamics are symmetric about the gravity axis, as in tilt recovery for multirotors, the value function cannot depend on heading. We then compute on the quotient: tilt direction on the sphere plus $\omega$, a 5-dimensional problem. A flat solver can only approximate this reduction by dropping yaw, and it would then still hit gimbal lock in pitch.
Proposed Experiments
No benchmark in the literature we compiled poses a controlled reach-avoid problem over the full rotation group, so we construct three, each tied to an existing system.
Benchmark 1 is kinematic attitude with rate as the control and a bounded rate disturbance. The state is $SO(3)$ alone, which is 3-dimensional. The failure set is a sensor keep-out cone, for example a star tracker that must not point near the sun. At this size a full grid on $SO(3)$ is feasible and serves as ground truth. We can also solve it with a grid-based reachability toolkit (helperOC or BEACLS, from Bansal et al.) on an Euler grid, which isolates the cost of gimbal lock with no learning involved.
Benchmark 2 is full rigid-body attitude on $SO(3)\times\mathbb{R}^3$, which is 6-dimensional, with torque bounds and gyroscopic coupling. The task is tumble recovery: reach a near-upright, low-rate target within a horizon from arbitrary orientations, while avoiding rates beyond actuator or structural limits. This benchmark echoes the 178° inverted recovery that the geometric $SE(3)$ controller of Lee et al. stabilizes. That paper proves stability and gives no reachable set.
Benchmark 3 is the heading-reduced 5D quadrotor tilt-recovery problem.
As a sanity check, we also run ConReach's own small-initial-set attitude forward-reachability problem. That paper reports 14.73 seconds for 40 time steps on a MacBook Air M2. We verify that our method's forward tube contains ConReach's ball where both apply.
The metrics are as follows. Volume error against ground truth, plus value error stratified by angular distance from Euler gimbal lock, on Benchmark 1. False-safe rate: the fraction of states declared recoverable that fail under the worst-case disturbance found by an adversarial disturbance optimizer during rollout. Certified fraction of the true safe set. Percentage of grid cells touched by the repair. Wall-clock and memory compared with a full $SO(3)\times\mathbb{R}^3$ grid, which we will attempt at coarse resolution only as a reference point.
Baselines
The naive reference is grid reachability in Euler angles with a tilt envelope. This is the near-hover practice used for quadrotor reachability in Bansal et al., here run with pitch restricted to below 60°. It is the "do nothing new" option because practitioners use it today. We run it on the same grid resolution budget and report both its accuracy inside the envelope and the fraction of the true safe set it cannot represent at all. Beating it means matching its accuracy inside the envelope while covering the full rotation group.
The strongest learned competitor is DeepReach with the same network size, training samples, and GPU-hours as our learner, under two orientation inputs: Euler angles, and unit quaternions with the value averaged over $q$ and $-q$. On a different problem, the 3D Air3D benchmark, DeepReach reports a mean squared error of $1.01\times10^{-4}$ and a volume error of 0.43% against a grid solution (Bansal & Tomlin). We have no reported result for any $SO(3)$ reach-avoid task, so we will establish it ourselves.
The key ablation is rotation-matrix-input DeepReach with no repair. This is our own learner alone, and it separates what geometric inputs buy from what certification buys.
The strongest certification competitor is HJ-PATCH run in Euler coordinates on its native Euclidean grid. It reported reducing raw neural-barrier safety failures from "up to 51%" to 0% on 2D and 4D systems (Tonkens & Herbert). We give it the same band-growth rule and node budget. Beating it means a lower false-safe rate near large tilts, at equal or fewer updated cells.
ConReach is included only on the small-set forward problem, because it does not compute controlled reach-avoid sets.
Our method counts as successful if three things hold. On Benchmark 1 it is within grid resolution of ground truth everywhere, while the Euler baselines show error concentrated near gimbal lock. On Benchmark 2 its false-safe rate is zero on grid-certified states, against nonzero rates for all unrepaired learners. And its repair touches a small single-digit percentage of cells, comparable to HJ-PATCH's reported 1.97% on its cruise-control example.
Why not just…?
Why not just restrict tilt and keep Euler angles? This works for hover-like flight, and our first baseline measures exactly that. It cannot certify tumble recovery, because the starting states of that task lie outside any singularity-free Euler box. Tomasi further shows that sensitivity spikes caused by the coordinates appear well before the singular angle. Benchmark 1, stratified by distance from gimbal lock, tests directly whether those artifacts degrade reachability accuracy inside a typical envelope.
Why not just feed rotation matrices or symmetrized quaternions into DeepReach? This is a real threat. As argued above, the rotation-matrix version already removes gimbal lock and double covering, which is why we adopt it. What it does not provide is a worst-case certificate: DeepReach's training is approximate and can settle in local minima. The ablation "rotation-matrix learner without repair", compared against the full pipeline on false-safe rate, settles whether the certified repair is worth its cost. If the unrepaired learner's false-safe rate is already zero under adversarial rollouts, the repair stage adds little beyond a formal statement. We will report that outcome as it is.
Why not just extend ConReach with multiple charts? ConReach's metric balls are tight for small sets but rely on strong convexity. Their authors note that joining charts across the whole rotation group is unresolved, and its output is a forward tube, not a controlled recovery set. The sanity check quantifies how much larger our tube is than ConReach's where both apply. Benchmark 2 shows the regime of whole-group reach-avoid problems that ConReach does not address.
Why not just switch Euler conventions near singularities? Switching conventions turns the problem into hybrid reachability with discrete jumps at chart boundaries. A single grid on $SO(3)$ avoids those jumps altogether. Whether the jumps matter in practice will show up as boundary error at chart seams, if we run a two-convention grid on Benchmark 1.
Decisive Checks
The idea fails if any of three things happen. First, Euler-grid errors on Benchmark 1 are not concentrated near gimbal lock and are no worse than our grid's errors, meaning the geometry does not matter in practice. Second, the repair band on Benchmark 2 grows to cover most of the grid, meaning the learned guess is too poor for local repair to escape the curse of dimensionality. Third, the unrepaired rotation-matrix learner matches the full pipeline's false-safe rate under adversarial disturbance search. In that case the honest conclusion is that geometric network inputs alone suffice, and certification is a formality.
References
- Geometric Reachability for Attitude Control Systems via Contraction Theory
- DeepReach: A Deep Learning Approach to High-Dimensional Reachability
- Parameterized Fast and Safe Tracking (FaSTrack) using Deepreach
- Exact and efficient Hamilton-Jacobi reachability for decoupled systems
- Patching Neural Barrier Functions Using Hamilton-Jacobi Reachability
- A Discrete Variational Integrator for Optimal Control Problems on SO(3)
- A Physics-Informed Machine Learning Framework for Safe and Optimal Control of Autonomous Systems
- Hamilton-Jacobi Reachability: A Brief Overview and Recent Advances
- Control of Complex Maneuvers for a Quadrotor UAV using Geometric Methods on SE(3)
- Representation Issues in the ML Estimation of Camera Motion