Cross-Embodiment Transfer

Cross-embodiment enables policies and demonstrations to be transferred across different robot morphologies.

See Robots Dataset for the full list of supported robots.

---

Basic Usage

Switching Robots for Tasks

For tabletop manipulation tasks with parallel grippers, you can swap robots easily:

# Run StackCube with Franka (default)
python scripts/advanced/replay_demo.py --sim=isaaclab --task=StackCube --num_envs=4

# Run the same task with KUKA iiwa
python scripts/advanced/replay_demo.py --sim=isaaclab --task=StackCube --num_envs=4 --robot=iiwa

# Run with UR10
python scripts/advanced/replay_demo.py --sim=isaaclab --task=StackCube --num_envs=4 --robot=ur10

---

Retarget between Robots

We provide src/scripts/retarget_demo.py to retarget trajectories from one source robot to one or multiple target robots.

Requirements

You need to go over:

• [Get Started / Installation / cuRobo Installation] for cuRobo

• Review Data format v2 for the data format

• Make sure that the following items are carefully set in the robots’ meta configs:

  • gripper_open_q / gripper_close_q: A list specifying the gripper’s joint positions when it releases / grasps the object

  • curobo_ref_cfg_name: cuRobo config file for the robot

  • curobo_tcp_rel_pos / curobo_tcp_rel_rot: Relative transformation from the TCP frame to the EE frame

    • The “EE frame” here is the ee_link specified by the cuRobo config

@configclass
class BaseRobotMetaCfg(ArticulationObjMetaCfg):
    # ...

    gripper_open_q: list[float] = MISSING
    gripper_close_q: list[float] = MISSING

    # cuRobo Configs
    curobo_ref_cfg_name: str = MISSING
    curobo_tcp_rel_pos: tuple[float, float, float] = MISSING
    curobo_tcp_rel_rot: tuple[float, float, float] = MISSING

Source Data and Configurations Preparation

To perform cross-embodiment retarget, you need to get robot configurations for the source and all the target robots prepared. You also need a demo data (.pkl) that contains the trajectory.

The robot meta config should include the information about the Tool Center Point (TCP) frame: On which link’s frame is it defined, and the relative transformation. Ideally, if the TCP link is already defined, you can

Retarget

python src/scripts/retarget_demo.py --source_path data_isaaclab/source_data/maniskill2/rigid_body/PickCube-v0/trajectory-unified_v2.pkl --source_robot franka --target_robots iiwa franka_with_gripper_extension

The exported pickle file with contain the original demos as well as the retargetted demo for the target robots (see Data format v2):

{
    "franka": [  // robot name should be same as BaseRobotMetaCfg.name
        // Demo for Franka
        "actions": [ ... ],
        "init_state": { ... },
        "states": [ ... ]
        "extra": ...
	],
	"iiwa": [
       ... // Demo for KUKA IIWA
	],
	"ur10": [
       ... // Demo for UR10
	]
}

By specifying --viz, the first retargetted trajectory will be visualized via plotly in your browser.