Humanoid robots and the tasks their specifications actually describe

The same robot name can cover different hardware

Unitree's G1 specification separates the standard G1 from G1 EDU. It lists 23 joints for the standard model and 23 to 43 for EDU configurations. Optional hand and wrist hardware contributes to that difference. The table also lists secondary development for EDU and warns that maximum arm load changes with extension posture. [1]

Consequently, a research video using a G1 with additional hands does not establish what ships with the base model. Keep the hand, computing module, software access and mechanical version alongside the name. Avoid treating a maximum joint count as a count of independently useful manipulation actions.

Scroll sideways for all columns.

Walking and handling an object interact

Use a tray-carrying task as an engineering thought experiment. As the arms extend, the carried mass moves farther from the body. The controller must account for the changed load while keeping the tray in the intended orientation. A walking clip without the tray leaves this condition untested.

Break the task into approach, grasp, lift, travel, placement and release. During each stage, observe the object's state as well as the robot's posture. A robot remaining upright does not establish that the contents stayed in the tray.

For a comparison, request trials at the required object heights and carry distances. Record whether the path was mapped, whether an operator issued movement commands and whether the robot had external support. These conditions prevent a locomotion result from being reused as an unsupported manipulation claim.

What BMW's Figure 02 trial establishes

In its 2024 report, BMW described Figure 02 placing sheet-metal parts into fixtures at Spartanburg. The fixtures were used for a later body-production step. BMW discussed evaluating how the robot connected with production systems. This is evidence for a named trial on a defined part-handling job. [2]

The report is not a complete production benchmark. It does not provide all attempts, intervention counts or a cost per accepted part. Reading the trial this way preserves the documented task without inventing a broader deployment result.

Score the complete placement rather than the grasp

Consider an illustrative 24-trial evaluation across three object positions. The robot lifts the part in 22 trials, places it correctly without assistance in 18 and needs an operator in four of the completed lifts. The grasp rate is 22/24, about 91.7%. Unassisted correct placement is 18/24, or 75%.

This gap identifies where more evidence is needed. Check transport, alignment and release before assuming the gripper is the only limitation. Keep the results separated by starting position; one accessible pose should not hide a consistently unsuccessful pose.

Include setup and recovery in the time log. The example is not a measured result for Figure, Unitree or another manufacturer.

Compare a humanoid with a simpler installation

For a fixed pickup-and-placement station, request an equivalent proposal using a fixed arm. For a task that spans several stations on a suitable floor, include a wheeled mobile manipulator. Compare accepted placements, site changes, operator involvement and service requirements on the same task.

A humanoid proposal needs evidence for the parts of the job that justify its legs and body. That might be access geometry or movement between working positions, but the benefit must be shown for the actual site. There is no universal ranking that follows from a human-like shape.