Robotic Planning Under Environmental and Perceptual Uncertainty

41 concepts40 papers1 proposal

When a robot operates in a simulated environment, planning is straightforward. The simulator provides exact coordinates, physics are perfectly predictable, and actions execute flawlessly. This idealized perspective relies on the closed-world assumption—the premise that the robot possesses complete, static knowledge of its environment and that actions yield deterministic outcomes.

In the real world, this assumption shatters. Actuators slip on uneven terrain, sensors encounter noise and occlusions, and environments change dynamically. If a robot relies on classical, deterministic planning, a single unexpected obstacle or a slight mechanical misalignment can cause catastrophic failure.

To build truly autonomous systems, roboticists must design planning frameworks that treat uncertainty not as an exceptional error, but as a core characteristic of the environment.

The Limits of Deterministic Execution

Classical motion planning algorithms typically focus on finding the shortest or most computationally efficient path from a start state to a goal state. However, when physical execution is subject to stochastic disturbances—such as wind gusts hitting a drone or wheels slipping on sand—the mathematically optimal path can become hazardous.

To address this, researchers shift the planning objective from minimizing distance to maximizing the probability of success. For example, Alterovitz et al. introduced the Stochastic Motion Roadmap (SMR). By integrating sampling-based roadmaps with Markov Decision Processes (MDPs), this framework explicitly models motion noise during the planning phase. Instead of assuming a robot will perfectly follow a trajectory, the algorithm calculates the likelihood of deviation, steering the robot along paths that are inherently more forgiving of mechanical slippage.

[Deterministic Planning]  Start ───────► Shortest Path (High Risk) ───────► Goal (Fails if robot slips)
                                              ▲ (Obstacle Nearby)

[Stochastic Planning]     Start ───────► Safer, Wider Path ───────────────► Goal (Succeeds despite noise)

As systems grow more complex, accounting for these low-level physical dynamics alongside high-level goals becomes computationally expensive. Shoukry et al. resolved this by decoupling high-level pathfinding from low-level dynamics using a "lazy" Satisfiability Modulo Theories (SMT) approach. This allows the planner to maintain strict safety guarantees in reach-avoid problems even when dealing with high-dimensional state spaces subject to external disturbances.

In fast-moving scenarios, a robot cannot afford to pause and re-plan when a disturbance occurs. The authors of Robust and Recursively Feasible Real-Time Trajectory Planning in Unknown Environments designed a real-time trajectory planner that pre-computes a sequence of invariant tracking "funnels." These funnels form a closed loop of safe states, ensuring that even if a robot is buffeted by external forces, it is guaranteed to remain within a safe, feasible operating envelope without needing to stop. Similarly, Zeng et al. incorporated probabilistic uncertainty directly into the trajectory logic of autonomous vehicles, ensuring smooth, human-like adjustments when handling hardware sensor noise during urban lane changes.

Uncertainty does not only stem from imperfect execution; it also arises from imperfect perception. A robot's sensors provide a noisy, partial view of the world.

Traditional computer vision systems often rely on binary classification—identifying an object definitively as X or Y. In non-ideal environmental conditions, this approach is brittle. To mitigate this, Luan et al. proposed a framework utilizing high predictive value classifiers that allow a robot to express its degree of uncertainty or defer a decision entirely, preventing downstream planning failures caused by false recognitions.

When interacting with physical objects, a robot rarely knows their exact geometric boundaries. Gaschler et al. bridged the gap between high-level symbolic task planning and low-level geometric motion planning through their KABouM framework. By representing objects and robots as bounding sets of convex polyhedra, the system allows the robot to reason about tasks safely and perform rapid collision checking without requiring perfect geometric data.

[Exact Geometry (Brittle)]   Robot Arm ───► Attempts precise grasp ───► Misses due to 1mm sensor error
[Bounding Polyhedra (Robust)] Robot Arm ───► Reasons over volume bounds ──► Successful, error-tolerant grasp

This perceptual uncertainty becomes even more pronounced in multi-robot systems, where localization errors compound across the fleet. Sung et al. demonstrated that moving away from rigid, synchronous execution models allows multi-robot teams to handle localization and object-pose uncertainty asynchronously. By decoupling the robots' timelines, a single sensor error or delay does not cause the entire multi-robot plan to stall.

Active Knowledge Acquisition and Contingencies

When faced with an uncertain environment, a passive robot must guess or wait. An active robot, however, treats information gathering as an explicit goal.

Consider a robot encountering a closed container: it cannot know if the container is heavy or light simply by looking at it. To handle such hidden internal states, Gaschler et al. developed the Knowledge of Volumes (KVP) framework, which generates contingency plans based on run-time sensing. If the robot lacks information, it schedules a sensing action to resolve the ambiguity before proceeding.

Taking this concept further, Chen et al. formalized physical actions, sensing, and human dialogue into a single logic-based language (Action Language $BC$). If a service robot cannot locate an object due to an incomplete world model, its generated plan will explicitly include actions like "asking a human for help" or "searching the next room."

Is object location known?
   ├── Yes ──► Execute direct manipulation path
   └── No  ──► [Active Knowledge Acquisition] ──► Ask human / Look around ──► Update model ──► Execute

When navigating environments containing movable obstacles, the physical outcomes of interacting with those obstacles are often unpredictable. Zhang et al. modeled observation, model, action, and partial observability as time-cost intervals. This allows a robot to make risk-aware decisions on the fly, calculating whether it is faster and safer to navigate around an obstacle or attempt to push it out of the way.

To handle highly dynamic environments where obstacle trajectories are unknown, modern frameworks frequently combine classical global search with learned local behaviors. Angulo et al. paired global pathfinding with a learned local steering policy, enabling robots to navigate safely around dozens of dynamic obstacles with unpredictable movements. For aerial delivery systems facing environmental uncertainty like wind, Chen et al. reformulated non-convex temporal logic and collision constraints into smooth, recursive convex forms, enabling high-quality trajectory adjustments in fractions of a second.

This integration of execution, planning, and real-time learning mirrors foundational insights from cognitive architectures. As Laird et al. demonstrated within the Soar architecture, an agent operating under noisy perception must be capable of seamlessly transitioning between high-speed reactive reflexes and deep deliberate planning when it encounters an environmental impasse. When an action fails, the system must diagnose why it failed—whether due to a physical constraint or an impossible task—and adapt its parameters rather than blindly repeating the same mistake (Karapinar et al.; Kiesel).

Conclusion

Robotic planning under environmental and perceptual uncertainty represents a paradigm shift from finding the optimal path in a known world to finding the safest policy in an unpredictable one. By explicitly modeling motion noise, embracing geometric approximations, and treating knowledge acquisition as a plannable action, modern robotic frameworks successfully bridge the gap between abstract symbolic reasoning and the messy reality of the physical world.

Key Foundational Work

Decision Analysis For An Experimental Robot With Unreliable Sensors (1975) — contributed 1 concept elements to this direction.
Application of Theorem Proving to Problem Solving (1969) — 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_environmental_uncertainty_5511c8a5,
  title = {environmental uncertainty},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/environmental-uncertainty/rdr_5511c8a5db7e439f93aaa7e795d3a0dc},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_environmental_uncertainty_5511c8a5_v1,
  title = {environmental uncertainty},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/environmental-uncertainty/rdr_5511c8a5db7e439f93aaa7e795d3a0dc/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake