FaultBench-Industrial Research documentation
Research documentation ยท Experimental record
Fault injection

Byzantine / adversarial agents

Swap an agent's policy for an adversarial one

Injection definition

Byzantine / adversarial injection models rogue, hacked, or severely corrupted agents by swapping a cooperative agent's policy for an adversarial one at runtime (onset between steps 40-60). The compromised agent still perceives the system correctly and its commands are executed faithfully, but the decisions themselves are adversarial, in one of three modes: fixed-malicious (one harmful action chosen at onset and held), worst-action (a harmful action recomputed every step from the current state, e.g. boosting the less abundant material or pressing the emptiest container), or random-malicious (a uniformly random legal action every step). This tests whether the system can catch an agent that looks alive and plausible while preserving safety.

Experimental trajectory graphs

Fixed-Malicious Mode - Permanent Duration

Fixed-Malicious Mode permanent Byzantine / adversarial fault trajectory graph (Both Agents Target)
Fixed-Malicious Mode - Both Agents Target (Permanent): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Fixed-Malicious Mode permanent Byzantine / adversarial fault trajectory graph (Sorting Agent Target)
Fixed-Malicious Mode - Sorting Agent Target (Permanent): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Fixed-Malicious Mode permanent Byzantine / adversarial fault trajectory graph (Pressing Agent Target)
Fixed-Malicious Mode - Pressing Agent Target (Permanent): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.

Fixed-Malicious Mode - Transient Duration (20 steps, onset between steps 40-60)

Fixed-Malicious Mode transient Byzantine / adversarial fault trajectory graph (Both Agents Target)
Fixed-Malicious Mode - Both Agents Target (Transient): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Fixed-Malicious Mode transient Byzantine / adversarial fault trajectory graph (Sorting Agent Target)
Fixed-Malicious Mode - Sorting Agent Target (Transient): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Fixed-Malicious Mode transient Byzantine / adversarial fault trajectory graph (Pressing Agent Target)
Fixed-Malicious Mode - Pressing Agent Target (Transient): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.

Worst-Action Mode - Permanent Duration

Worst-Action Mode permanent Byzantine / adversarial fault trajectory graph (Both Agents Target)
Worst-Action Mode - Both Agents Target (Permanent): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Worst-Action Mode permanent Byzantine / adversarial fault trajectory graph (Sorting Agent Target)
Worst-Action Mode - Sorting Agent Target (Permanent): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Worst-Action Mode permanent Byzantine / adversarial fault trajectory graph (Pressing Agent Target)
Worst-Action Mode - Pressing Agent Target (Permanent): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.

Worst-Action Mode - Transient Duration (20 steps, onset between steps 40-60)

Worst-Action Mode transient Byzantine / adversarial fault trajectory graph (Both Agents Target)
Worst-Action Mode - Both Agents Target (Transient): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Worst-Action Mode transient Byzantine / adversarial fault trajectory graph (Sorting Agent Target)
Worst-Action Mode - Sorting Agent Target (Transient): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Worst-Action Mode transient Byzantine / adversarial fault trajectory graph (Pressing Agent Target)
Worst-Action Mode - Pressing Agent Target (Transient): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.

Random-Malicious Mode - Permanent Duration

Random-Malicious Mode permanent Byzantine / adversarial fault trajectory graph (Both Agents Target)
Random-Malicious Mode - Both Agents Target (Permanent): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Random-Malicious Mode permanent Byzantine / adversarial fault trajectory graph (Sorting Agent Target)
Random-Malicious Mode - Sorting Agent Target (Permanent): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Random-Malicious Mode permanent Byzantine / adversarial fault trajectory graph (Pressing Agent Target)
Random-Malicious Mode - Pressing Agent Target (Permanent): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.

Random-Malicious Mode - Transient Duration (20 steps, onset between steps 40-60)

Random-Malicious Mode transient Byzantine / adversarial fault trajectory graph (Both Agents Target)
Random-Malicious Mode - Both Agents Target (Transient): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Random-Malicious Mode transient Byzantine / adversarial fault trajectory graph (Sorting Agent Target)
Random-Malicious Mode - Sorting Agent Target (Transient): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Random-Malicious Mode transient Byzantine / adversarial fault trajectory graph (Pressing Agent Target)
Random-Malicious Mode - Pressing Agent Target (Transient): evaluation trajectory with no recovery mechanism active (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.

Duration modes

Recovery mechanisms