Planning Safe Trajectories within the Sensor Horizon

40 concepts28 papers1 proposal

When an autonomous robot navigates an unfamiliar environment, it must constantly make assumptions about what lies beyond its immediate line of sight. Traditional motion planners often operate under an optimistic heuristic: they treat unobserved or unmapped regions as free space until a sensor explicitly detects an obstacle (Beul et al.).

While this optimism allows for fast, fluid trajectories in sparse environments, it introduces severe risks in cluttered or dynamic settings. A drone zipping around a blind corner or a self-driving car turning past a large building can suffer from a "corner surprise"—committing to a high-speed path through a blind spot where a hazard might suddenly appear (Zhou et al.). If the robot's speed exceeds its ability to react, it enters an Inevitable Collision State (ICS): a physical configuration from which no future control actions can prevent a crash (Janson et al.).

To guarantee safety without global map data, a growing body of research focuses on constraining motion planning based on the robot's immediate visibility. The core intuition is simple: never enter a state where a collision with an unobserved obstacle is physically unavoidable.

The Geometry of Visibility-Aware Planning

To mathematically guarantee safety in partially observable environments, a robot's physical movement must be strictly coupled with its perceptual field of view. Goretkin et al. formalize this constraint by requiring that a robot's swept volume—the total physical space its body occupies during a maneuver—never intersect with unverified or occluded regions.

[ Visible Free Space ]  -->  Robot can safely execute swept volume
----------------------
[ Visibility Boundary ] -->  Treated as a hard obstacle / source of threats
----------------------
[ Unobserved Space ]    -->  Assumed occupied or hazardous

This principle transforms the motion planning problem. Instead of planning a path purely toward a destination, the robot's state representation combines its physical configuration with the cumulative history of its observed environment.

To maintain this safety boundary, a vehicle's velocity must be directly constrained by its current sensor visibility horizon (Tas). As demonstrated by Mitsch et al., ground robots can maintain formal safety guarantees by ensuring they never travel into a blind area at a speed higher than what is required to come to a complete stop within the currently visible region. If a sensor's field of view shrinks—due to an obstruction, a sharp turn, or sensor range limits—the vehicle must automatically reduce its speed or widen its trajectory to preserve a safe reaction distance (Zhou et al.).

Treating the Unknown as an Obstacle

How do planners enforce these visibility constraints computationally? The most direct method is to treat the boundary of unsensed space as a physical obstacle within the safety framework.

By "wrapping" around existing path planners, algorithms can treat all unmapped or occluded regions as occupied space (Bronstein et al.; Zhou et al.). This forces the optimization engine to generate trajectories that always maintain a safe exit path staying entirely within the currently visible region (Janson et al.).

This conservative mapping manifests across different robotic domains:

  • 2.5D Terrain Mapping: When navigating rugged landscapes, robots assign a conservative maximum height upper bound to unobserved cells, preventing the robot from stepping onto overhanging structures or hidden drops until they are explicitly scanned (Miki et al.).
  • 3D Wire-Frame Reconstruction: In mobile vision systems, projecting a plane from the sensor's viewpoint through a single detected edge provides a conservative estimate of occupied space in occluded regions, ensuring the robot does not assume hidden areas are traversable (Utsumi et al.).
  • Mechanical Manipulation: When searching for objects on cluttered shelves, maintaining a conservative existence map that treats occluded regions as hidden obstacles prevents erratic or overconfident gripper movements (Kim et al.).

Worst-Case Occlusion Modeling

In dynamic environments, such as urban intersections, treating unobserved space as a static wall is insufficient. A blind spot could hide a pedestrian or another vehicle moving at high speed.

To account for this, researchers model the worst-case scenario within the unobserved space. For instance, Kurzer et al. propose a heuristic where occluded regions of a road are treated as if they contain a virtual vehicle traveling at the maximum allowed speed limit. Similarly, Tas places hypothetical obstacles right at the boundary of the sensor's visible range.

This worst-case modeling directly alters the robot's behavior. Because the mathematical optimization requires a valid stopping path to exist for any potential hidden threat, the vehicle exhibits emergent, human-like cautious behaviors. It will proactively slow down or "creep" forward at blind intersections, accelerating only once its sensors have swept the unobserved space and verified that the path is clear (Yu et al.; Tas).

Balancing Conservatism and Efficiency

Strictly adhering to worst-case assumptions can lead to the "freezing robot" problem, where a system refuses to move because it cannot prove that an unobserved region is completely safe (Elhafsi et al.). To maintain high performance while preserving safety, modern frameworks employ reachability analysis and active perception.

Bidirectional Reachability

Instead of assuming threats exist everywhere in the unknown space, Yu et al. combine forward reachability (where the ego vehicle can physically travel) with backward reachability (where a hidden threat must originate to cause a collision). By filtering potential risks against real-time visibility data, the system ignores blind spots that are mathematically incapable of producing a colliding threat, allowing the vehicle to maintain higher average speeds.

Dynamic Safety Horizons

When dealing with communication lags or intermittent sensor data, tracking the reachable sets of obstacles rather than their last-known positions ensures the robot maintains a robust safety buffer (Cleaveland et al.). Vaskov et al. derive a formal minimum sensor horizon—a metric that defines the exact range a sensor must cover based on the robot's braking capabilities and the maximum velocity of surrounding obstacles to guarantee a fail-safe maneuver can always be executed.

Active Verification

Rather than guessing what lies in a blind spot, planners can actively schedule movements to maximize visibility. Stephens et al. identify intermediate "progression candidates"—locations that offer a better vantage point to inspect high-value but uncertain areas. By moving to these positions first, the robot gathers the necessary visual data to verify safety before committing to a path. This active verification prevents robots from taking inefficient or risky detours through narrow passages that are likely to be blocked by unseen obstacles (Zhang et al.; the authors of Enabling Topological Planning with Monocular Vision).

Key Foundational Work

Inevitable collision states. A step towards safer robots? (2003) — contributed 3 concept elements to this direction.
Estimating Surface and Spatial Structure from Wire-Frame Model Using Geometrical & Heuristical Relation (1992) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_conservative_occlusion_modeling_cddf8942,
  title = {conservative occlusion modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/conservative-occlusion-modeling/rdr_cddf89429b05463f961793310ac7dadb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_conservative_occlusion_modeling_cddf8942_v1,
  title = {conservative occlusion modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/conservative-occlusion-modeling/rdr_cddf89429b05463f961793310ac7dadb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake