A robot that builds a 3D map from its camera while it moves has to decide, every time the image disagrees with the map, what caused the disagreement. There are three common causes. The scene may contain static geometry the map has not yet resolved. Something may be moving. Or the surface may be shiny, so its appearance changes with viewpoint. These explanations are often observationally degenerate: several different physical situations produce the same pixels. Current mapping methods settle the question by fiat and keep one answer. Current safety filters then treat that answer as ground truth. When the chosen answer is wrong, the robot either drives into a glass wall that was dismissed as a "transient reflection," or freezes in front of a ghost obstacle left behind by a pedestrian who has already walked away.
We propose to stop guessing. The safety filter should carry the full set of explanations that the camera data cannot yet exclude, and it should enforce collision avoidance against all of them. The robot can also shrink that set cheaply by choosing small, informative motions. A brief stop separates true motion from parallax, and a sidestep separates parallax from specular sliding. This borrows the physicist's way of breaking degeneracy, which is to add a probe whose ambiguity runs in a different direction. The result should be a navigation filter that is provably no less safe than one that keeps every suspicious primitive as an obstacle. It should also make far more progress, because ambiguities get resolved by moving rather than by waiting or freezing.
Core Claim
Our central claim is about where safety failures in vision-based navigation through dynamic scenes come from. Most of them come not from noise in the map but from structured ambiguity, meaning discrete, mutually exclusive explanations of the same residual. Two properties of this ambiguity matter. First, it can be represented as a small set of hypotheses for each ambiguous map region. Second, the robot's own motion can break it, without the intrusive interventions that physical-science degeneracy breaking normally needs. A barrier condition that is robust over the retained hypothesis set, with a calibrated rule for when a hypothesis may be dropped, gives a finite-sample probabilistic safety guarantee. It does so with much less conservatism than uniform margin inflation.
To make this concrete, we use the explicit Gaussian map of SAFER-Splat. That work builds a control barrier function (a scalar "distance-to-unsafe" function whose rate of decrease is bounded, so that a quadratic program can minimally correct the robot's commanded action) directly on the ellipsoids of an online 3D Gaussian Splatting map. It runs at 15 Hz and explicitly assumes a static world.
For each map region $j$ whose photometric residual is unexplained, we maintain a retained set $R_{j,t}$ drawn from three kinds of hypothesis. Under the static hypothesis, the region is occupied and fixed. Under the dynamic hypothesis, it is occupied and moving with some velocity $v$ in a bounded cone $V_j$. Under the virtual hypothesis, the appearance does not correspond to occupied space, as with a reflection or a shadow. The filter enforces
$$\min_{\eta \in R_{j,t}} \left[ \dot h_j^{\eta}(x,u) + \alpha\, h_j^{\eta}(x) \right] \ge 0 \quad \text{for all } j,$$
where $x$ is the robot state and $u$ its control. The term $h_j^{\eta}$ is the robot-to-region distance barrier under hypothesis $\eta$; for dynamic hypotheses its time derivative includes the obstacle velocity. The constant $\alpha > 0$ sets how fast the barrier may shrink. Virtual hypotheses contribute no constraint, but they can only win if every occupied hypothesis has been excluded.
This structure follows the logic of two earlier works. Barrier Certificates for Unknown Systems with Latent States builds one certificate valid across an entire posterior ensemble of degenerate dynamics models. Shielding under Partial Observability overrides any action unsafe for some state in the belief support. Our setting is different from both: the hypotheses are spatial scene explanations rather than ODE parameters or grid states, and the filter must run at camera rate.
Research Program
The work has three parts.
Hypothesis generation. We extend an online Gaussian mapper with a per-region ambiguity detector. The detector fires where residuals are persistent or where densification is triggered by error spikes, which is the mechanism RobustSplat identifies as the source of mid-air floaters. Each flagged region gets the three hypothesis families. The dynamic family is parameterized by per-Gaussian velocities in the style of Split4D. We reuse Split4D's velocity-carrying primitives but not its commitment to a single decomposition.
Calibrated pruning. A hypothesis is removed only when a likelihood-ratio test against the competing hypotheses exceeds a threshold. We set that threshold by conformal calibration (a distribution-free way of choosing thresholds from held-out labeled data so that the error rate is at most a chosen $\epsilon$). Calibration uses held-out sequences with ground-truth labels for static, moving, and virtual regions, so that the true explanation is wrongly pruned with probability at most $\epsilon$.
This is the step where ReCORS supplies the method. ReCORS calibrates dynamics-model error with conformal prediction under full state observability. We instead calibrate a perception-level exclusion test, which is exactly the observability gap ReCORS leaves open. The guarantee inherits conformal prediction's exchangeability assumption between calibration and deployment scenes. We will state this assumption plainly and test it under deliberate distribution shift.
Active degeneracy breaking. This part is motivated by Li et al., who separated water absorption from liner reflectivity by changing photon step length, and by Cosmic Complementarity, which sums Fisher information from probes whose degenerate directions cross. Each candidate motion primitive predicts a different residual signature under each hypothesis. Under a stop, static and virtual regions stay fixed in the image while moving ones drift. Under a lateral sidestep, static geometry shows parallax that matches its depth, whereas a mirror reflection slides according to the mirror's geometry. Under a rotation in place, the hypotheses separate along yet another axis.
A small planner scores each primitive by its expected reduction in retained-set size. It then chooses between the nominal action and a probe, subject to the barrier constraint above. Probing is therefore always safe, because it is itself filtered.
The analysis component will characterize when the robust quadratic program becomes infeasible. This can happen when a robot is surrounded by unexcluded dynamic hypotheses. We will show that, in that case, the stop primitive is the maximally informative probe for separating dynamic from static hypotheses, which gives a principled fallback.
Baselines
The do-nothing reference is SAFER-Splat as published: a static online Gaussian map, a double-integrator barrier, and no dynamic handling. It establishes how badly the static assumption fails once moving objects and reflective surfaces are introduced. No reported result was found on Lacuna for SAFER-Splat in dynamic scenes. The only reported safety figure is static-scene drone flight with "only minor, brief violations of the safety condition (on the order of 1-4mm)," attributed to controller tracking error.
The conservative reference keeps every flagged primitive as a static obstacle and never prunes. By construction it is the safest static-geometry filter against standing obstacles. It is the bar our method must match on collisions with static and glass obstacles while beating it on progress.
The strongest existing competitor is a point-estimate dynamic pipeline. We take Split4D's decomposition, which reports 0.893 mIoU on Multi-Human against 0.592 for the tracking-based SA4D, and feed its static Gaussians and velocity-carrying dynamic Gaussians into a velocity-aware version of the SAFER-Splat barrier. This is the obvious construction at this intersection. It is strong because Split4D's decomposition is the best reported in the file. We will also run a variant that uses RobustSplat's transient masks for the static layer. RobustSplat is reported to beat SpotLessSplats, WildGaussians, and T-3DGS in PSNR, SSIM, and LPIPS, although no raw numbers were found on Lacuna.
The calibrated-margin baseline wraps that same point-estimate pipeline in conformal inflation. It inflates every obstacle by a radius calibrated on the same held-out data so that occupancy coverage reaches $1-\epsilon$.
All methods share the same mapper backbone, the same camera streams, the same calibration sequences, the same robot dynamics, and the same per-method tuning budget of a fixed number of configurations searched on validation scenes. Each runs within the same real-time budget of at least 15 Hz on identical hardware. Our method counts as beating the strongest competitor only if it does two things. It must show a lower collision rate in scenes with contact zones and specular surfaces. It must also show no worse task success than the calibrated-margin baseline at the same $\epsilon$.
Three ablations isolate the ingredients. The first removes active probing and keeps the retained set passively, which isolates complementarity. The second replaces the retained set by its most likely hypothesis, which isolates set-keeping. The third uses an uncalibrated fixed pruning threshold, which isolates the conformal step.
Why not just…?
Why not just keep every suspicious primitive as a static obstacle? This is safe against things that stand still, and it is a real competitor on collision rate in scenes with only static obstacles. It fails in two ways. Moving obstacles are represented by stale positions with no velocity, so the barrier does not anticipate an approaching pedestrian. And RobustSplat's analysis of densification-induced floaters implies that ghost obstacles accumulate along every path a distractor took, which should cause freezing. Our experiments report freezing rate and time-to-goal against this baseline directly. If it matches us on success in our scenes, then the probing mechanism is not earning its cost.
Why not just use the best dynamic reconstruction and add obstacle velocity to the barrier? Because a point estimate is wrong exactly where it matters. Split4D itself reports shared primitives and blurred boundaries in contact zones. RobustSplat and NeRF On-the-go both acknowledge that specular and illumination changes get misclassified as transient. For a robot, that misclassification deletes a real wall. The decisive test is the per-category collision breakdown in the glass and contact-zone scenes described below.
Why not just inflate margins with conformal prediction around a single map? This is the most serious threat. It is simpler and comes with the same kind of guarantee. Our argument is that the ambiguity is bimodal: an obstacle is either there or not, and either moving or not. A single inflation radius that covers both modes must be large everywhere, so it should over-inflate unambiguous geometry. We will test this directly by comparing narrow-passage success at matched $\epsilon$. If uniform inflation matches us, we will report that the hypothesis structure is unnecessary in these regimes.
Why not mask moving objects with a pretrained segmenter, or learn safety end to end? Split4D's 0.593 versus 0.893 mIoU comparison shows that tracking-based masks degrade under occlusion and interaction. Masking also discards the motion of the very obstacles to be avoided. For end-to-end learning, Safety-Gymnasium reports that PPO-Lagrangian cut violations by 98% at the cost of a 45% reward drop. ReCORS reports that Lagrangian baselines could not verify safety beyond 20 steps, against more than 50 for ReCORS. We include a segmentation-mask variant but not a learned policy, since the question here concerns certified filters.
Proposed Experiments
Benchmark construction. No benchmark in the file combines dynamic scenes with certified navigation, so we will build one. We extend the five simulated SAFER-Splat navigation scenes, which contain 15K to 500K Gaussians, in three ways: Kubric-style controlled insertion of moving objects, mirror and glass surfaces, and objects that touch static surfaces. Ground-truth occupancy and per-region explanation labels are available by construction. Calibration and test splits use disjoint scenes. A shifted test split adds surface materials unseen during calibration, to probe the exchangeability assumption.
Navigation metrics. We measure collision rate, broken down by obstacle category (static, moving, glass, contact). We also report minimum signed distance, task success, time-to-goal, freezing rate, quadratic-program infeasibility rate, and control frequency.
Perception-level validation. Separately, we validate the hypothesis tests on real data. Using the RobustNeRF benchmark and the On-the-go sequences, we check whether the empirical rate of wrongly pruned true explanations stays at or below $\epsilon$. For reference, RobustNeRF reports 30.87 dB PSNR against 25.22 dB for Mip-NeRF 360, and NeRF On-the-go reports 21.41 dB on Patio against 20.54 dB for RobustNeRF. These sequences are not robot-controlled, so the probing component cannot be tested on them. We will instead select sub-sequences whose camera motion resembles a stop or a sidestep, and measure how much each kind of motion shrinks the retained set. Contact-zone behaviour is checked on Multi-Human against Split4D's segmentation.
Decisive checks. Three outcomes would decide the proposal. First, empirical miss rate tracks $\epsilon$ on calibration-matched scenes, and we measure how much it degrades under shift. Second, on glass and contact-zone scenes, our method's collision rate is at or below the conservative baseline, while the point-estimate dynamic pipeline is strictly worse. Third, at matched $\epsilon$, our method beats calibrated uniform inflation on narrow-passage success, and the no-probing ablation recovers most of that gap only at the cost of longer time-to-goal. If the third check fails, the idea reduces to conformal inflation and we will say so.
References
- A Control Barrier Function for Safe Navigation with Online Gaussian Splatting Maps
- Barrier Certificates for Unknown Systems with Latent States and Polynomial Dynamics using Bayesian Inference
- Safe Reinforcement Learning via Shielding under Partial Observability
- RobustSplat: Decoupling Densification and Dynamics for Transient-Free 3DGS
- Split4D: Decomposed 4D Scene Reconstruction Without Video Segmentation
- Conformal Reachability for Safe Control in Unknown Environments
- Novel methods for measuring the optical parameters of the water Cherenkov detector
- Cosmic Complementarity: Joint Parameter Estimation from Cosmic Microwave Background Experiments and Redshift Surveys
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild
- Safety Gymnasium: A Unified Safe Reinforcement Learning Benchmark