Models for the Python SDK

​

SO-101

observation = dreamscale.so101.observe(
    side_frame=side,
    wrist_frame=wrist,
    joint_positions=joints,  # 6 values
)

• Joint order: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper.
• Joints are in calibrated LeRobot degrees; the gripper is 0–100.
• Actions are absolute targets in the same order and units.
• flux-3-action-so101 builds on the actions it already sent, so call it with next_action, not predict.

​

Bimanual YAM

observation = dreamscale.yam.observe(
    top_frame=top,
    left_frame=left_wrist,
    right_frame=right_wrist,
    left_joint_positions=left_joints,    # 6 values, radians
    left_gripper=left_gripper,           # 0 closed to 1 open
    right_joint_positions=right_joints,  # 6 values, radians
    right_gripper=right_gripper,         # 0 closed to 1 open
)

• Actions are 14 absolute targets: left joints 1–6, left gripper, right joints 1–6, right gripper. Grippers are at indices 6 and 13.
• Grippers are continuous. Don’t threshold them to open or closed.

​

LIBERO

observation = dreamscale.libero.observe(
    agent_frame=agent_view,  # agentview_image
    wrist_frame=wrist_view,  # robot0_eye_in_hand_image
    state=state,             # 8 values
)

• State: end-effector position (3), end-effector rotation as an axis-angle vector (3), then the two gripper joint positions.
• Pass frames as LIBERO renders them. The SDK flips them upright.
• Actions are LIBERO deltas: dx, dy, dz, three axis-angle rotation deltas, then the gripper (−1 open, +1 close).