Sensor noise

Corrupt observations

Injection definition

Sensor noise introduces stochastic or deterministic corruptions (such as Gaussian noise, biased offsets, or zero-out corruptions) directly into the local observation vectors of targeted agents. During injection onset (randomized between environment steps 40–60), the agent receives miscalibrated sensory feedback about belt composition, sorting purity, and container fill states while the physical environment continues to operate normally. This tests how effectively agents can maintain optimal control under distorted observation perception.

Experimental trajectory graphs

Permanent fault mode

Permanent sensor noise fault trajectory graph — Permanent sensor noise fault trajectory showing the active fault window (onset at step 40, persisting through the evaluation episode), impacted sorting purity, container fill levels, and cumulative reward degradation without recovery intervention.

Transient fault mode (40–60 timesteps)

Transient sensor noise fault trajectory graph (40-60 timesteps) — Transient sensor noise fault trajectory showing the bounded fault window active between steps 40–60. Displays sorting purity recovery, container fill levels, 16767 total material processed, and final cumulative reward of 71.75 without recovery intervention.

Duration modes

Permanent: the fault persists for the relevant evaluation.
Transient: the fault is introduced for a bounded interval.

Recovery mechanisms

Rule-based reconfiguration
Fault-tolerant-trained MARL
Hierarchical supervisor-agent
LLM-agent replanning