srb agent ros — ROS 2 Agent

Control a simulation through ROS 2 middleware.

Usage

srb agent ros --env ENV_ID [options]

Options | Shared Agent Options

Without --ros_sync, the latest available action is sampled each step and the loop is paced to env.agent_rate. Slower publishers may therefore reuse an action. --ros_unthrottled removes pacing; without env.agent_rate, the loop is unpaced. When wall-clock pacing is enabled, a configured env.agent_rate must be finite and positive. Action subscriptions are best-effort, keep-last depth one.

Commands expire after 0.5 seconds of monotonic wall time by default. Fresh messages renew only their addressed environment and action slices; velocity and event inputs expire independently. Malformed or non-finite messages do not refresh commands. Mapped actions must be finite, are clamped to the batched Box bounds, and those bounds must admit zero. Raw/term overrides own their slices until expiry or reset. The CLI wait timeout is separate from expiry; Python embedders may set RosInterface(env, command_timeout_s=0.25) or disable expiry with None. Velocity and event traffic updates only mapped or unclaimed slices; a fully overridden mapped update is not counted as applied.

Reset, expiry and action publication are synchronized. Contended reads return zero. The agent holds action_transaction() across env.step(action); Python embedders sharing the interface with reset or close callbacks should do the same and must not retain the borrowed action buffer after the context exits. Nested reads are rejected. A hung consumer can delay reset; this is not a watchdog. A busy close retains ingress and dependent endpoints for retry. Host interleaving checks do not qualify native ROS/CUDA behavior.

Zero action is not a controller watchdog, emergency stop or proof of physical safety. Hardware use requires separate transport and physical-stop acceptance.

RealEnv raw policy input

srb real_agent ros --env ENV_ID accepts flat policy actions on action/raw (normally /srb/action/raw) as std_msgs/Float64MultiArray, in the generated environment’s single-action order and bounds. RealEnv applies its hardware mapping and scales. The real adapter uses this topic; simulator term, velocity and event mappings do not define real teleoperation semantics.

The subscription is best-effort, volatile, keep-last depth one. Valid first input must arrive within 5 seconds; commands expire after 0.5 seconds, checked before dispatch and each prepared hardware command. A positive finite --ros_action_timeout overrides both deadlines. --ros_sync waits for a newer message after the first step; if the previous message expires first, the session aborts. --ros_unthrottled does not bypass expiry or RealEnv’s action rate.

Missing, malformed, out-of-bounds or expired commands abort the real session; invalid input is not replaced with zero. Cleanup stops the session, closes the command source and RealEnv, then releases the process-owned ROS context. Failed cleanup retains resources for retry. Message age is measured from local callback receipt because messages have no producer timestamp. These checks cannot interrupt a blocked hardware call or provide a watchdog; synthetic transport checks do not establish physical-stop acceptance.

See the ROS 2 Workflow for examples.