Amazon researchers are testing a Unitree G1 on tasks useful in a warehouse: picking up boxes, carrying them and opening doors. The humanoid robot has to coordinate its arms, legs and balance. In the video, it carries a box while ducking under a low beam.
This is FIBER, a research project involving Amazon FAR, Carnegie Mellon University and Stanford University. The trials take place in a test environment. The team is exploring how a single controller can combine different tasks; the footage does not show a robot already working in an Amazon warehouse.
The arms cannot work alone
Lifting a box changes body posture and weight distribution. On the FIBER project page, real trials include boxes in different positions and heavy doors opened by leaning into them. Waypoints can guide the robot through whole-body motions while carrying, including crouching under an obstacle.
Skills are trained separately in simulation and brought into a shared behaviour model. A planner then chooses commands using depth information and the state of the body. It replans five times a second while the lower-level controller adjusts fifty times a second. Choosing the next movement need not run at the same pace as correcting balance.
From a trick to a sequence
The useful step is making locomotion part of manipulation. The robot changes its posture to suit the task rather than treating walking and handling as unrelated abilities.
This remains research. The reported average of 42 percent success across three simulated tasks is not a reliability rating for a delivery service. The physical demonstrations instead show what becomes possible when a humanoid uses its whole body as one coordinated system.
