Skip to main content

Execution and Concurrency

Three kinds of parallelism are independent: multiple robots in one scene, multiple simulated environments, and multiple models inside an agent system.

Selecting robots​

InitializationParameters.agent_names supplies public robot names. Their order must match the environment's robot order. With the default ['default'], initialization resolves the actual names. Observation.selected_robot_name identifies the robot for a tool call; use it with get_agent in multi-robot tools.

Shared resources​

execution_group groups trajectories that must run sequentially within one execution batch. The converter inserts delays in call order for tools sharing the group. It does not provide a global lock across episodes or replan trajectories after each preceding tool finishes.

ToolBatchContext supports reservations while tools in the same request are prepared, such as reserving different placement cells. Reservations and trajectory groups solve different problems: choosing distinct resources and ordering their use.

Preserving unrelated actors​

ToolExecution.allowed_moving_actors declares which actor changes an executor may preserve. In the current GEN executors, other actors are restored from the saved source state after tool execution.

  • None: no explicit protection.
  • []: no actor may retain changes.
  • A list of actor names: those actors may retain changes.

For multiple calls, allowlists are combined; any None disables batch protection. is_detection=True forces the tool allowlist to [].

This mechanism restores actor state; it does not guarantee collision-free motion. It applies to compatible actors state entries, not all articulations or arbitrary world data. Invalid names or incompatible layouts cause the executors to skip protection with a warning.

Next: Coordinate robots. Exact fields: tool reference.