Action Space Catalog

Each robot offers one or more action modes, chosen at launch via env.robot.action_mode=<mode> (the default mode is marked *). Dims gives action-vector length; Scale multiplies policy output before it reaches the controller. runtime-resolved length depends on the loaded asset (e.g. a joint-space mode with pattern .*) and is read from the action manager when the environment is built.

The machine-readable action spec (env.srb_spec.action) targets the action manager’s runtime terms. Schema-version-2 descriptors name runtime terms directly, e.g. the multicopter cmd_vel, body-acceleration body_vel and task-space delta_twist terms. A schema-version-1 descriptor may split one runtime term into semantic fields, such as lin and ang for the wheeled-drive cmd_vel term, but its summed field dimensions must equal the summed runtime term dimensions. A mismatch of either kind fails the spec build instead of publishing a wrong layout. Groups without descriptors use the schema-version-0 fallback: no semantics, no dimension check. Metadata that carries an unsupported descriptor (the multicopter thrust / multicopter.rotor_thrusts, the task-space quaternion ee_rot / end_effector.orientation, or the velocity-labelled body lin / body.linear_velocity) is rejected on load.

Task-space modes publish one runtime-length delta_twist term: three position-delta channels (m), then three axis-angle rotation-delta channels (rad); position-only inverse kinematics uses the first three. Variable-stiffness operational-space modes append six stiffness channels, and full-impedance modes append six stiffness and then six damping-ratio channels.

Multicopter Body Acceleration (body_acc on the multicopters) is a velocity command despite its name: four channels (linear_x, linear_y, linear_z, angular_z) are clamped to [-1, 1] and multiplied by the scale into m/s and rad/s, and x and y are rotated by the current yaw. Body Acceleration (body_acc on the orbital bodies) is an acceleration command: six channels are clamped to [-1, 1] and multiplied by the scale into a linear (m/s²) and angular (rad/s²) acceleration in the world frame, which is integrated into the root velocity as acceleration * physics_dt on every physics step. The velocity change over one environment step is therefore acceleration * step_dt whatever the decimation. The relative controller variant, when configured, expresses the same command in the body frame.

Task-space modes (diff_ik, osc_*, smooth_osc) act on the tool centre point that tasks observe as robot_ee (the tf_end_effector target): the manipulator’s flange frame composed with the mounted tool’s TCP, or the bare flange when no tool is mounted. The environment config sets every such term’s body_offset to that frame. Isaac Lab’s task-space terms carry the body Jacobian to it in the articulation root frame, the frame the end-effector pose and velocity are measured in (a term’s base_name, when set, must name the root body), and the operational-space terms damp the TCP velocity derived from the flange link frame rather than its centre of mass. A pure rotation command therefore turns the tool about its TCP.