Humanoid Robot Testing Standards

Claims such as works in a warehouse are too broad. A test should define the object mass shelf height walking distance allowed retries time limit human assistance and what counts as completion. The result becomes meaningful because another team can reproduce the same conditions.

Capability asks whether the robot can do the task at all. Robustness asks whether it can keep doing it when object pose lighting floor friction or starting position changes. Both numbers matter. A high success rate in one fixed setup can hide poor generalization.

Perception planning control locomotion manipulation and communication should be measured separately and together. Component tests help localize faults. End to end task tests expose interactions between layers such as a planner that creates motions the controller can execute only near the center of its stability margin.

Stale data, missed grasps and lost connectivity are failure cases to consider in a controlled test plan. Fault injection on physical hardware requires a risk assessment, protection measures and defined stop conditions. The cited project notices do not authorise arbitrary fault injection or provide a completed humanoid safety protocol.

Report robot version software build control mode sensors test date sample size and environment. State when teleoperation or human reset was used. A benchmark becomes useful when the reader can see both the headline score and the conditions that produced it.