DemoHLM shows generalizable loco-manipulation on Unitree G1 from a single simulated demonstration

Key points

• The DemoHLM paper presents a hierarchical loco-manipulation framework that learns multi-task behavior from a single demonstration in simulation instead of large real-world training sets.
• The authors report real-world validation on a Unitree G1 robot with an RGB-D camera, covering ten loco-manipulation tasks under spatial variation.
• It matters for G1 because it positions the platform as a repeat-use research testbed for generalizable whole-body manipulation rather than only a locomotion demo robot.