Skip to main content

Coordinate Robots and Shared Resources

First make each tool work on one robot. Then specify robot identity and how simultaneous calls share resources.

Robot identity​

Configure InitializationParameters.agent_names in environment order. The runtime injects known_robots and verifies its consistency. In a tool, select the robot with self.get_agent(obs.selected_robot_name) instead of assuming the first robot.

Serialize a shared workspace​

For a fixed resource, add execution_group="shared_tray" to the decorator. For a call-dependent resource, return ToolExecution(..., execution_group=target_zone).

Calls sharing the group are delayed in call order within that execution batch. Different groups may proceed concurrently. This is not a global lock, and trajectories are not automatically replanned after preceding group members finish.

Reserve placement locations​

Tools in one request share obs.tool_batch_context. During preparation:

reserved = obs.tool_batch_context.try_reserve(
    namespace="placement_cells",
    resource=("tray", 0, 0),
    owner=obs.selected_robot_name,
)

On failure, choose another cell or return a failed execution. The same owner can reserve the same resource again. Reservations live for the current request, not for the whole episode; use world state to represent persistent occupancy.

The installed PlacementGrid utility combines geometric placement choices with these reservations. Consult its API before assuming a grid cell is free just because no object was observed there before the batch started.

Preserve only intended actor changes​

Return allowed_moving_actors=[object_name] for an object manipulation if that matches your simulation policy. Every simultaneous call needs an explicit list for the batch protection to remain enabled. Detection tools use is_detection=True, which forces an empty list.

This restoration policy is separate from shared-zone scheduling. It does not make a movement physically safe. See execution concepts.

Validate concurrency​

The manual tool CLI accepts one call at a time; it is useful for each robot individually. Then run your preset with GEN to verify simultaneous calls, reservations, group delays, and interruption of running skills.