Workspace-Anchored Local Codes for Neural Reachability

Before a robot can be trusted near obstacles, someone has to compute the set of states from which it can always avoid collision, even under worst-case disturbances. This "safe set" is the zero level of a value function that solves a Hamilton-Jacobi partial differential equation, which describes how safety propagates backward in time under the system's dynamics. Grid solvers cannot store this function beyond about six state dimensions. Neural solvers such as DeepReach replace the grid with one sinusoidal network trained on the equation itself. These networks scale to 9 and 10 dimensions, but they blur exactly the parts of the safe-set boundary that matter: the dents and corridors carved by thin or closely spaced obstacles. In 3D shape modeling the same blurring was fixed by giving each small region of space its own latent code, as in Local Implicit Grid Representations for 3D Scenes. The obvious transfer fails, though. A local grid over a 10-dimensional state space is just the dimensionality problem again.

The proposal rests on an observation about where the fine detail lives. In most reachability problems, the sharp structure comes from obstacle geometry, and obstacle geometry is defined over a few workspace coordinates: planar or spatial position, or pairwise relative position between agents. The remaining coordinates, such as velocity, heading, and other agents' internal states, change the value function smoothly. We therefore place a local latent grid only over the low-dimensional workspace coordinates and query it at the state's projected position. A shared decoder then combines the retrieved code with the full state and time. Memory grows with workspace resolution raised to two or three, not with the state dimension. The local codes can store obstacle-induced detail that a single global network smooths away.

Core Claim

Neural reachability solvers lose safe volume in obstacle-rich, high-dimensional problems because one global network must represent workspace-localized, high-frequency boundary structure. Anchoring local latent codes to workspace coordinates recovers that structure with memory that does not depend on state dimension. Concretely, we write the value function using the exact-boundary form of Imposing Exact Safety Specifications in Neural Reachable Tubes:

$$V_\theta(x,t) = l(x) + (T-t)\, D_\theta\big(\mathrm{interp}(Z, p(x)),\, x,\, t\big)$$

The symbols are as follows. $l(x)$ is the signed distance to the failure set. $T$ is the horizon. $p(x)$ maps the state to its workspace coordinates, of dimension $k \in \{2,3\}$. $Z$ is a learnable grid of latent codes over the workspace, with each code a 32-dimensional vector as in the local implicit grid work. $\mathrm{interp}$ is multilinear interpolation over the $2^k$ neighboring cells. $D_\theta$ is a small sinusoidal decoder. The decoder is not a deep monolithic network; its job is to carry the smooth dependence on dynamics. For multi-agent problems, $p$ returns each pair's relative position. The same grid is queried once per pair, and the codes are pooled symmetrically, so the grid is shared across pairs instead of growing with the number of agents.

The approach makes a falsifiable prediction. Gains should concentrate on problems whose safe-set boundary has workspace-localized detail, such as the 5D bicycle with five obstacles, rocket landing on a constrained pad, and the 10D narrow passage. On Air3D, whose boundary is smooth, gains should be negligible. If the architecture helps uniformly, our story about why it works is wrong, even if the numbers look good.

Relation to Prior Work

Two papers supply the method. The local implicit grid work showed that interpolated per-cell codes beat a global latent on unseen shapes, with mean F-score 0.692 against 0.274. We reuse its interpolated-code design but move the grid off the full domain and onto a projection. The exact-boundary paper hard-codes $l(x)$ so the terminal condition holds by construction. We keep that form and change only the network inside it, which remains a monolithic sinusoidal network in that paper. Progressive Surface Reconstruction from Images Using a Local Prior motivates the work, because it showed that global smoothness priors shrink thin structures. It also offers a blending alternative, Tukey-weighted patch fusion, which we test against plain interpolation.

Two grid-based papers constrain the design, because they show the ceiling we are trying to get around. Refining Control Barrier Functions through Hamilton-Jacobi Reachability warm-starts global dynamic programming from a candidate barrier and stops at six dimensions because of memory. HJ-PATCH updates only boundary states that violate safety, reaching 1.97% of states updated on the 2D adaptive cruise control task, yet it still allocates the full grid. Our method is not a grid solver and makes no convergence claim of the kind these papers prove. In six dimensions and below we use HJ-PATCH as a reference and as a repair step.

C-IRIS also decomposes space into regions, but its regions are kinematic polytopes without dynamics. That paper shares our intuition that workspace geometry should drive the decomposition. Blending Neural Control Density Functions proves a union law for blended local controllers. That is a stronger kind of composition than ours, but it has been verified only up to 6D.

Research Program

We first instantiate the architecture on the reachability benchmarks reported for DeepReach and ExactBC: Air3D, 5D bicycle, 6D rocket landing, 9D multi-aircraft, and 10D narrow passage. Training uses the exact-boundary loss and time curriculum from ExactBC, applied unchanged. The grid codes and decoder are trained jointly on the equation residual, with no autoencoder pretraining, since there are no shape crops to pretrain on.

Second, we characterize the method's limits. We sweep workspace grid resolution and code size against recovered volume. We also design a stress case in which the sharp structure lives partly in non-workspace coordinates. Heading-dependent narrow gaps for the bicycle are one such case. These cases show when the projection assumption breaks.

Third, the network gives no guarantee. To get a safety-relevant output, we calibrate the zero level with the statistical approach of Conformal Reachability for Safe Control in Unknown Environments, shrinking the level set until a held-out rollout violation rate meets a target. In six dimensions and below, we run HJ-PATCH starting from our network. This tests whether a sharper starting point means fewer patched states, which would be a concrete benefit for formal repair.

Baselines

The naive reference sets $V(x,t)=l(x)$, treating the obstacle signed distance as the value. It shows how much of the safe set any learned method must carve away, and its violation rate shows why dynamics matter.

The strongest existing method is ExactBC. The file reports safe-volume recovery of 61.1% on the 5D bicycle and 33.4% on 6D rocket landing, against DeepReach's 41.4% and 0.4% (Imposing Exact Safety Specifications in Neural Reachable Tubes). For the 9D and 10D tasks, no reported ExactBC result was found on Lacuna, so we will rerun it. DeepReach is included with its reported Air3D figures, mean squared error $1.01\times10^{-4}$ and volume error 0.43%. Both are rerun from their protocols on identical sampled states, the same GPU-hour budget, and an equal hyperparameter search budget. Our model's total parameters, including the code grid, are matched to the baseline's.

Beating ExactBC means higher recovered safe volume at an equal or lower conformally calibrated violation rate on the bicycle and rocket tasks. The target is a clear margin above 61.1% and 33.4%. Matching or exceeding ExactBC on the 10D passage at equal compute also counts. On Air3D we expect parity, so we will not claim improvement there.

In six dimensions and below, HJ-PATCH and grid solvers serve as ground truth and as a formal reference. The comparison metrics there are recovered volume and patched-state count.

The key ablation keeps everything fixed but moves the code grid onto a random 2D linear projection of the state instead of the workspace. If this projection does as well, the gain comes from extra parameters, not geometric anchoring. A second ablation removes the codes and gives the parameters to the decoder. A third uses a coarse full-state local grid at equal memory, to show directly how the obvious transfer degrades as dimension grows.

Why not just…?

Why not just raise SIREN capacity or frequency, or add Fourier features? This is the most serious threat, because spectral bias is partly a question of frequency scale. The evidence against it is indirect. DeepReach's collapse to 0.4% on rocket landing came largely from loss imbalance, which ExactBC fixed. The residual 33.4% suggests a further representational limit, but does not prove one. We settle this with a parameter-matched sweep over SIREN frequency scale and Fourier-feature bandwidth on the bicycle and rocket tasks. If a tuned global network closes the gap, we will report that the local codes are unnecessary.

Why not just use adaptive grids or octrees with dynamic programming? Adaptive refinement still subdivides every state axis. HJ-PATCH and REFINECBF both stop at six dimensions despite pruning, because the grid must be allocated. Our grid never sees the non-workspace axes.

Why not just use a local grid or multiresolution hash encoding over the full state? A dense grid over 10 dimensions is exponential in size. Hashing it causes collisions that mix unrelated distant states. The coarse full-state-grid ablation and a hash-encoding run at matched memory test this directly.

Why not just patch ExactBC with HJ-PATCH? This works only in six dimensions and below, where we do run it. Beyond six dimensions it has no grid to patch on, which is exactly the regime we target.

Proposed Experiments

The decisive check combines three conditions. On the bicycle and rocket tasks, recovered volume must exceed ExactBC's reported 61.1% and 33.4% at matched calibrated violation rates. The random-projection ablation must fall back toward ExactBC. The Air3D gain must be near zero. Together these confirm that the improvement comes from workspace anchoring.

On 9D multi-aircraft and 10D narrow passage there is no grid ground truth. There we report rollout violation rates under worst-case disturbances, sampled-residual maps near obstacles, and recovered volume estimated by Monte Carlo rollout checks.

In six dimensions and below, a reduction in states that HJ-PATCH must patch, starting from our network versus from ExactBC, would show that sharper neural boundaries also make formal repair cheaper.