G1 MP (AMO)

Whole-body G1 tasks whose demonstrations are produced by automated motion planning. They run on LocoManipulationEnv with the g1_wholebody robot, whose lower body is driven by the AMO policy (robots/policy/AMO_Policy.py, amo_jit.pt): CuRobo plans the dual-arm kinematic trajectory while a scripted policy issues the base locomotion commands.

Collect data with the Motion Planning pipeline:

uv run datagen simple/G1WholebodyBendPickMP-v0 \
  --sim-mode=mujoco_isaac --render-hz=50 --no-headless --num-episodes=10

Environments (5)

  • simple/G1WholebodyBendPickAndPlaceOnSofaMP-v0

  • simple/G1WholebodyBendPickMP-v0

  • simple/G1WholebodyPickAndBendPlaceMP-v0

  • simple/G1WholebodyTabletopGraspMP-v0

  • simple/G1WholebodyTabletopHandoverMP-v0

The fixed-base Franka, Aloha, Vega, G1 and G1-Inspire tabletop environments also use the motion-planning pipeline; see Franka Tabletop Grasp and Aloha Tabletop Grasp.