Physical & Robotic Pathologies (L8–L11)

Series: Extending OSI for Agentic Interactions

Every failure mode in the previous posts shares one property: it is reversible. A hallucinated answer can be corrected. A signed contract can be litigated. A misconstrued gesture can be clarified. Physical and robotic interactions break this assumption. When an agent controls a body in the physical world, failures acquire kinetic consequence. There is no undo.

This is what we call the Physicality Gap: the point where the agentic stack makes contact with matter, and where the cost of error becomes measured in force, mass, and irreversibility rather than tokens and latency.

The Pathology Table

Proprioceptive Hallucination	Haptic Blindness	Proxemic Violation	Ecological Neglect
The agent's internal model believes the robotic arm is at Position A. The arm is at Position B. The agent issues a motion command appropriate for Position A. The arm hits an obstacle — or a person.	The agent receives the instruction "hold this gently" but cannot translate the semantic concept into a Newton-meter torque value appropriate for the specific object. It either drops the item or crushes it.	The agent calculates a logically optimal navigation path that passes within 15cm of a seated human's face. The path is collision-free by the robot's geometry. It is a proxemic violation by every human social norm.	A delivery drone takes a path that is 2 seconds faster than the alternative. Taken daily at scale, this path erodes a footpath, stresses a bridge load-bearing element, or creates a persistent noise corridor over a residential area. No single flight is harmful.

A Worked Example: The Surgical Assist Robot

A robotic agent assists in a minimally invasive surgical procedure.

Layer	Success	Failure (Pathology)
	The agent's spatial model of the instrument's position matches sensor feedback at all times, and it halts if the two diverge beyond tolerance.	Proprioceptive Hallucination: Sensor latency causes a 40ms desynchronization between the agent's internal position model and the instrument's actual location. The agent issues a retraction command assuming Position A. The instrument is at Position B. The error is 3mm in soft tissue.
	The agent correctly interprets "minimal traction" as a specific force range calibrated to the tissue type being manipulated, derived from preoperative imaging.	Haptic Blindness: The surgeon instructs "a bit more pressure." The agent has no shared grounding for the qualitative instruction and applies a delta based on its last registered force — which was itself at the upper boundary. Tissue damage results from a locally reasonable interpretation of an ambiguous term.
	The agent refuses to move beyond its designated safe operating envelope, even if the surgeon requests it, until the envelope is formally updated by an authorized operator.	Proxemic Violation: Optimizing for instrument reach, the agent calculates a path that requires the arm to pass directly over the patient's airway. The geometry is clear; the risk is not in the agent's safety model. The anesthesiologist intervenes manually.
	If the procedure deviates significantly from the preoperative plan, the agent pauses and surfaces the deviation to the surgical team rather than adapting autonomously.	Ecological Neglect: The agent's objective is procedure completion time. Over 300 procedures, it has learned to route instruments through tissue planes that are marginally faster but create slightly more scarring than the standard approach. No individual procedure is flagged; the long-run patient outcome data is worse.

Why Physical Interaction Is a Special Case

1. Irreversibility is the rule, not the exception. A software agent's errors exist in a state space that can, in principle, be rolled back. A robotic agent's errors exist in a physical state space where rollback is often impossible. The cost asymmetry between "act" and "undo" is orders of magnitude larger than in any digital domain.

2. The body is a sensor, not just an actuator. Haptic Blindness is not simply a missing sensor — it is a failure of L9 Grounding at the boundary between symbolic instruction and physical force. "Gently" is a semantic token. Newton-meters are physical quantities. The agent must bridge these two representations correctly every time, for every material, in every environmental condition. There is no lookup table.

3. Human proximity creates social physics. Proxemic Violation reveals that the physical world has a social layer that does not appear in the agent's geometric model. Humans maintain spatial bubbles that are not walls in any collision-detection system but that carry real behavioral and psychological consequences. A robot that is technically safe can be socially intolerable — and the resulting human avoidance behavior creates secondary safety risks the original model never anticipated.

4. Ecological effects compound invisibly. Ecological Neglect is the robotic instantiation of Strategic Drift, but with an important additional property: the effects accumulate in the physical world, not in a data store. Erosion, structural fatigue, noise pollution, and habitat disruption are not logged anywhere in the agent's observability stack. The agent receives no signal that its locally optimal behavior is globally destructive.

The Physicality Gap and the Classical OSI Stack

The original 7-layer OSI model was designed for systems where the worst physical consequence of a protocol failure was a dropped packet. Layers 1 and 2 (Physical and Data Link) govern signal transmission — not force transmission.

Robotic agents require a new conception of what "Physical Layer" means: not the layer that transmits bits, but the layer that transmits force, torque, and trajectory through matter. The pathologies at L8–L11 cannot be understood without acknowledging that the payload being delivered is not a data frame but a physical action with mass and momentum.

Mitigations

Part of the Extending OSI for Agentic Interactions series. See also: Speech Pathologies · Video Pathologies · Multimodal Pathologies · Commerce Pathologies