monitored-subtask-execution

Monitored Subtask Execution

Overview

Use a standardized procedure to call corobot_mcp_server MCP tools and execute one monitored robot subtask. In practice this is one CoRobot PolicyTask run defined by a prompt, an optional policy host/port, and an optional step_interval, followed by deterministic monitoring and cleanup.

Inputs (provide per run)

• prompt: task instruction to execute; keep it directly executable and unambiguous
• policy.host / policy.port: optional policy server; if omitted, the service default 127.0.0.1:8001 is used
• step_interval: optional step interval; service default is 1.5
• timeout_s: maximum time to wait for this run
• poll_interval_s: interval for polling get_status, for example 0.5 to 2.0
• reset_after: whether to call reset_task after completion and return to the home pose; collection workflows usually want true
• max_retries: number of retries after failure; usually 0 to 2

Tools (MCP)

Core tools, named as {service_name}___{tool_name}:

• corobot_mcp_server___set_evaluate_params: set the prompt, policy, and step_interval, then auto-start the task after a short delay
• corobot_mcp_server___get_status: retrieve PolicyTask status to determine running, success, or failure state
• corobot_mcp_server___stop_task: stop the current task; prefer this before manual intervention
• corobot_mcp_server___reset_task: stop and reset the robot to the initial pose

Optional tools, usually not needed:

• corobot_mcp_server___start_task: explicit start; usually unnecessary because set_evaluate_params already auto-starts
• corobot_mcp_server___get_prompt / corobot_mcp_server___set_prompt: inspect or set the current prompt

For exact argument structure, see references/mcp-tool-map.md.

Quick Actions

Hard reset (no prompt)

When no prompt should be executed and you only need to return the robot to a reusable start state:

1. Call corobot_mcp_server___stop_task if a task might still be running.
2. Call corobot_mcp_server___reset_task.

Workflow: Execute One Subtask (single prompt)

Step 0: Safety + Preconditions

• Confirm the robot workspace is safe, emergency stop is available, and human intervention is possible.
• Confirm the MCP service is enabled and reachable. corobot_mcp_server is configured in src/agent_demo/config/ormcp_services.json.
• Decide whether this run should preserve its terminal state or return to a reusable start state:
• If only the action result matters, use reset_after=false.
• If the next round needs a reusable start state, use reset_after=true.

Step 1: Start (set params + auto-start)

Call corobot_mcp_server___set_evaluate_params.

Argument template:

{
  "evaluate_params": {
    "policy": {"host": "127.0.0.1", "port": 8001},
    "prompt": "<prompt>",
    "step_interval": 1.5
  }
}

Important behavior:

• On success, set_evaluate_params waits briefly and auto-calls start_task. Do not call start_task again unless you have a specific reason.

Step 2: Monitor (poll get_status with a timeout)

• Poll corobot_mcp_server___get_status every poll_interval_s until one of these conditions is met:
• the returned status clearly indicates completion, success, or failure
• timeout_s is reached
• behavior becomes unsafe or manual intervention is required
• Save the raw status text from each poll as a log record even if it cannot be parsed into a structured schema.

Step 3: Handle success / failure deterministically

Success path:

• If reset_after=true, call corobot_mcp_server___reset_task.
• Record the run metadata: prompt, policy host/port, step_interval, start time, end time, and final status text.

Failure, timeout, or uncertain-status path:

1. Call corobot_mcp_server___stop_task.
2. Call corobot_mcp_server___reset_task.
3. Record the failure reason, including timeout, connection failure, malformed status, or manual intervention.
4. If max_retries allows another attempt, return to Step 1. Otherwise mark the run as failed and exit.

Troubleshooting

• If the result indicates a CoRobot connection failure or request failure, verify that the local CoRobot HTTP service is reachable at http://localhost:8765 and that the robot control stack is healthy.
• If the task gets stuck, follow the failure path stop_task -> reset_task. Pause and restore the environment manually if needed.

References

• references/mcp-tool-map.md: corobot_mcp_server tool list, naming rules, and argument structure