Debugging and FAQ

Common Questions

1. Low grasp success rate — is it a calibration issue?

Start by inspecting the force/torque trace from the execution result (episode.observations.force_torque). An abnormal force spike at the moment of contact usually indicates TCP calibration drift or a force-sensor zero-point offset — re-run the end-effector calibration flow (TCP offset calibration + force-sensor zero calibration).

2. TaskGraph gets stuck at a node?

Check each node's status in the task-graph run summary. In most cases, the stuck node has no on_failure retry or branch configured — once it fails, there is no exit path and the graph hangs. Configure an explicit retry count or failure branch for critical nodes.

3. Data is not being uploaded?

Check whether the data pipeline (data_pipeline) upload mode was accidentally set to manual, and whether the network layer can reach the platform's data endpoint. Switching back to automatic upload mode restores the data flow.

4. Existing task scripts error out after replacing the end-effector?

Check whether fields such as grasp_mode and dof in the task description are compatible with the new end-effector profile — different end effectors support different grasp modes. Declare the supported grasp_modes explicitly in the end-effector profile, and add a compatibility check at the script level.