Research documentation
ยท
Experimental record
Research
Experiment design
Factor structure
| Factor | Levels |
|---|---|
| Fault type | 5 classes |
| Fault target | Sort agent, press agent, both (communication loss: press only) |
| Fault mode | 1-3 per fault type (e.g. stuck, restrict, slip for actuator degradation) |
| Fault duration | Permanent, transient |
| Recovery mechanism | 4 paradigms, run as 6 arms: no-recovery control, rule-based (oracle), rule-based with self-detection, supervisor, fault-tolerant MARL, LLM replanning |
| Evaluation | 8 random seeds per combination; 3 training seeds for the two trained mechanisms (4,160 runs in total) |
Fault timing
Each fault starts at a random step between 40 and 60 of a 200-step episode. Transient faults last 20 steps; permanent faults last until the end of the episode.
Comparison principle
Every recovery mechanism is evaluated against the same fault conditions and scored with the same metrics.