Planners built to be safe around blind corners follow one rule. The robot must always be able to stop inside the region it has actually seen to be empty, and it must assume that anything hiding in unseen space could come out at maximum speed. The rule is sound on paper. Its weak point is the phrase "seen to be empty." Every such guarantee in the literature takes the visible region from clean raycasting. Real LiDAR does not deliver that. It returns nothing off glass and dark surfaces, which makes a wall look like open space. It scatters off dust, which creates small phantom obstacles. Its error grows with range and with how steeply the beam hits a surface. Phantoms cast phantom shadows, a conservative planner treats every shadow as a place where an attacker could be hiding, and the robot freezes. Missing returns do the opposite and quietly break the guarantee.
We propose to keep the conservative stopping rule and change what it is applied to. The binary visible region becomes a certified free region. Its boundary is shrunk by range- and angle-dependent margins, and those margins are calibrated against a domain-randomized family of sensor degradations. Domain randomization here means training or tuning across many randomly perturbed simulated sensors so the real one looks like another sample. We also prune occlusion shadows too narrow to hide an agent of known minimum size, and we plan from the state the robot will be in when the solver's answer actually takes effect. The claim to test is specific. On a real ground robot with an unmodified commodity LiDAR, this calibrated planner can be tuned entirely in simulation and transferred zero-shot. It should keep the near-zero collision rate of idealized conservative planners while cutting the freezing that noisy scans induce.
Core Claim
Existing conservative occlusion planners assume crisp visibility. OA-MPC builds forward reachable sets for unseen agents emerging from occlusion boundaries at maximum speed and enforces a terminal stopping constraint at 10 Hz. Janson et al. outer-bound inevitable collision states, meaning states from which no maneuver can avoid a crash, using emergency-braking trajectories. Tas places virtual obstacles on the visibility edge. Tas comes closest to handling noise, but the noise it models is Gaussian error in the robot's own localization and in detected objects, not corruption of the visibility boundary itself.
The research file's own gap analysis, drawing on Deep Local Trajectory Replanning and the BARN Challenge report, points out that no existing work models how perceptual noise reshapes that boundary. The same analysis argues that the planners consuming crisp boundaries are therefore exposed to two opposite failure modes. We know of no reported real-deployment study that measures either one.
The new element is to treat the visible-free boundary as a measured quantity with calibrated error, so that two things become true. First, the stopping guarantee holds with stated probability over the randomized sensor family. Second, phantom occlusions that cannot geometrically hide an agent are removed without weakening that guarantee.
Research Program
The planner is a 2D model predictive controller in the OA-MPC style. Model predictive control here means re-solving a short-horizon trajectory optimization at every control step. Its terminal constraint requires that the robot can brake to rest inside the free set. The program changes three inputs to that controller.
Certified free region. For each LiDAR ray $i$ with measured free length $\hat d_i$, measured range $r_i$ and incidence angle $\phi_i$, the certified free length is
$$\tilde d_i = \hat d_i - m_\delta(r_i, \phi_i).$$
Here $m_\delta$ is the $(1-\delta)$ quantile of the overshoot $\hat d_i - d_i^{\star}$, where $d_i^{\star}$ is the true free length. The quantile is estimated per range and incidence bin from rays rendered under a randomized sensor model. That model covers beam divergence, range bias, dropout, maximum-range returns off specular and dark surfaces, dust speckle, and fog attenuation, each with randomized severity.
Conformal-style quantile estimation gives a finite-sample coverage statement for the randomized distribution. Coverage for the real sensor is not guaranteed. That is the empirical question the hardware experiments answer.
Rays that return nothing are treated in one of two ways. If the randomized model gives them a non-negligible probability of being a specular miss, they count as unknown, not as free to maximum range.
Size-aware emergence sources. Reachable-set methods seed hidden agents along every occlusion boundary. We instead seed them only from occluded pockets that could contain a disk of the minimum agent width $w$. The width test uses the certified free returns on both sides of the pocket. A single-beam dust speckle at range $r$ shadows a wedge only about $r\,\theta_{\text{beam}}$ wide, which is centimeters at typical ranges. Such a wedge cannot hide a pedestrian or a vehicle. This step is what should remove phantom-induced freezing, and it is conservative conditional on the minimum-size assumption.
Latency-aware commitment. We measure the solver's runtime distribution under the same randomization. Each solve then plans from the predicted state at time $t + L_q$, where $L_q$ is a high quantile of solve time. Inflating the hidden-agent reachable sets by $v_{\max} L_q$ is a simplified form of the latency-in-reachability gap the file identifies. If a solve overruns $L_q$, the controller keeps executing the previous plan's tail, which by construction ends in a verified stop.
We expect this component to matter less than the other two. Recursive feasibility from stopping constraints already handles most overruns. We include it mainly so that the sensing-to-actuation delay that decided the BARN finals is accounted for rather than ignored.
All margin computations are lookup tables plus one pass over the scan. Both must fit in a few milliseconds so the controller keeps OA-MPC's 10 Hz loop on embedded compute.
Baselines
The do-nothing reference is an optimistic MPC that treats unobserved space as free, with a reactive emergency brake layered on top. It isolates what any occlusion reasoning buys. OA-MPC reports that the occlusion-unaware MPC "consistently resulted in collisions" with occluded dynamic agents. We expect the same here and include it only to anchor the scale.
The strongest existing method is OA-MPC run directly on raw noisy scans. It is the published planner most closely matched to our setting: dynamic hidden agents, reachable sets from occlusion boundaries, and real-time stopping constraints. We reimplement it in the same 2D stack, with the same dynamics, horizon, $v_{\max}$ and compute. We found no reported collision-rate or speed numbers for OA-MPC on Lacuna, so it must be rerun.
For static unknown mazes, the reference numbers come from two papers. Janson et al. report a 0% collision rate across 500 hallway mazes, at 0.5–1.5 s per decision. Map-Predictive Motion Planning reports completion times within 7–8% of an oracle that knows the full map, running at about 2 Hz on 30 unseen mazes, with clean sensing in both cases. To beat OA-MPC we must, under realistic noise, match its collision rate with idealized sensing while significantly reducing freeze time and completion time.
Two filtered variants of OA-MPC are the fair "engineering fix" competitors. The first fuses scans over time with log-odds occupancy filtering, where each cell's occupancy belief accumulates evidence across scans before the planner sees it. The second inflates the visible boundary by a single uniform margin. Both get the same simulation tuning budget we use for calibration, with the margin and filter constants chosen on the same randomized scenes.
Ablations remove one ingredient at a time. One replaces range- and angle-dependent margins with a single calibrated scalar margin. One restores emergence from every occlusion boundary, disabling the size filter. One sets $L_q$ to zero.
Why not just…?
Why not just inflate the visible boundary by a fixed buffer and slow down when visibility is short? This is the most serious threat, and the scalar-margin ablation exists to settle it. Heuristic buffers have a poor record against formal reachability. The state-lattice planner in Towards Provably Not-at-Fault Control had 7.6% and 17.2% at-fault collisions where reachability-based design had 0.0%. Still, a buffer tuned on the same randomized data could be competitive. Our bet is that LiDAR error grows with range and incidence angle, so any single margin is either too wide nearby, costing speed, or too narrow far away, risking collisions. If the scalar-margin ablation reaches the same speed-safety tradeoff, we will drop the range-dependent model and report that a calibrated scalar margin is enough.
Why not just filter the scans temporally and feed OA-MPC as is? Log-odds filtering suppresses speckle, but each scan of smoothing adds delay, and it does nothing about persistent errors such as glass returning maximum range. The file notes that filtering assumes uncorrelated noise and breaks down under structured failures. The filtered baseline will show whether this holds on our degradation families, especially specular surfaces and fog.
Why not just train an end-to-end policy with heavy domain randomization? The BARN record argues against relying on it. The deep RL entry placed first in simulation with a score of 0.2415 but completed only 2/9 physical trials, while a classical planner with latency compensation completed 8/9 (BARN Challenge). We use domain randomization only to calibrate a few interpretable margins. This keeps the stopping guarantee intact and makes failures diagnosable.
Proposed Experiments
Simulation. We test in three settings:
- Mazes: the procedural 25 m mazes of Map-Predictive Motion Planning and the hallway mazes of Janson et al..
- BARN worlds: 300 BARN-style cellular-automata worlds (BARN Challenge).
- Hidden-agent scenarios: maze and BARN layouts with pop-out agents that emerge from occlusions at $v_{\max}$ at adversarially chosen times.
Calibration and evaluation use disjoint degradation families. For example, we calibrate without fog or glass and test with them, so the result measures out-of-family transfer rather than in-distribution fit. Metrics are:
- collision rate;
- completion time relative to a full-map oracle;
- freeze time, defined as time below 0.1 m/s while a feasible path exists;
- empirical violation rate of the certified free region against ground truth, which checks that $\delta$ is honored;
- deadline overruns.
Hardware. We run a Clearpath Jackal on cardboard courses modeled on the BARN finals, adding glass panels and a dust or fog source. A dummy pedestrian on a rail is released from behind occluders. All tuning is done in simulation; nothing is tuned on hardware.
Decisive checks. The ablations decide which ingredients survive. If the size filter does not reduce freeze time under speckle, that component is not useful. If range-dependent margins do not beat the scalar margin on the speed-safety tradeoff, the extra modeling is not justified. The headline result is the hardware comparison against OA-MPC on raw scans. Success means zero hidden-agent collisions with clearly lower freeze time and completion time, with the real-sensor violation rate of the certified region reported against its simulated target $\delta$.
References
- OA-MPC: Occlusion-Aware MPC for Guaranteed Safe Robot Navigation With Unseen Dynamic Obstacles
- Safe Motion Planning in Unknown Environments: Optimality Benchmarks and Tractable Policies
- Limited Visibility and Uncertainty Aware Motion Planning for Automated Driving
- Deep Local Trajectory Replanning and Control for Robot Navigation
- Autonomous Ground Navigation in Highly Constrained Spaces: Lessons learned from The BARN Challenge at ICRA 2022
- Map-Predictive Motion Planning in Unknown Environments
- Towards Provably Not-At-Fault Control of Autonomous Robots in Arbitrary Dynamic Environments