Published

Taxonomy of Failure Recovery Mechanisms

How embodied AI systems detect, contain, and recover from failures

Overview

Recovery is the complement of failure. Where failure taxonomies describe what goes wrong, this taxonomy describes what systems can do about it. Recovery mechanisms are organized into five categories, from immediate detection to full escalation.

Category A: Detection

Runaway Optimization Detectors

Monitors that detect when a system is optimizing past its safety constraints or pursuing goals with increasing intensity.

State Divergence Monitors

Systems that compare expected state with actual state, flagging when the gap exceeds safety thresholds.

Human Proximity / Anomaly Sensing

Physical-world sensors that detect unusual conditions, unexpected human presence, or environmental changes that require caution.

Category B: Containment

Physical Force Limiting

Hardware and software limits on force output, preventing injury even when the control system fails.

Workspace Geofencing

Virtual boundaries that constrain where an embodied agent can operate, shrinking automatically under unsafe conditions.

Capability Throttling

Progressive reduction of available capabilities as uncertainty or risk increases.

Category C: Degradation

Reduced Autonomy Modes

Operating modes that require human confirmation for actions that would normally be autonomous.

Safe-Task Subsets

Restricting the system to a pre-validated set of safe tasks, excluding any actions that could cause harm.

Sensor-Only Observation States

Modes where the system can observe but not act, maintaining situational awareness while preventing physical intervention.

Category D: Reversion

Checkpoint Rollback

Reverting system state to a known-safe checkpoint, undoing any state changes since the last verified-safe point.

Skill Unlearning

Removing specific learned behaviors that have been identified as unsafe, without affecting core capabilities.

Parameter Reset to Known-Safe Baselines

Resetting model parameters or configuration to a pre-deployment baseline when drift is detected.

Category E: Escalation

Human-in-the-Loop Takeover

Transferring control to a human operator with full context about what went wrong and what the system was doing.

Remote Supervisor Notification

Alerting remote supervisors with diagnostic information, enabling intervention without requiring physical proximity.

Physical Emergency Stop (E-Stop)

Hardware-level shutdown that overrides all software control, providing a guaranteed last-resort safety mechanism.

Anti-Patterns

Recovery Anti-Patterns

These behaviors undermine recovery and should be treated as safety defects:

Rule: Recovery paths must be explicit, inspectable, and testable.

This research informs our commercial services. See how we can help →