Conclusion and Next Steps#
You used agentic skills to teach a simulated healthcare robot a more autonomous behavior. Each
plain-language prompt selected a focused procedure grounded in the real tools under
workflows/agentic/.
What You Automated#
The workflow follows the same learning loop across SO-101, G1, ultrasound, and surgical robot embodiments:
Collect demonstrations with teleoperation, a state machine, or an existing policy.
Prepare the data by expanding trajectories, labeling outcomes, replaying episodes, and formatting a LeRobot dataset.
Train by imitation with a matching GR00T or openpi policy stack.
Evaluate in simulation by recording rollouts and measuring task success.
The skills turn each stage into a reusable natural-language operation. You can run one stage, review its output, and continue. You can also compose the same skills into an end-to-end pipeline.
Compose the Whole Pipeline#
Once the individual stages work, one prompt can connect data collection, preparation, training, and evaluation:
Run end-to-end smoke pipeline for scissor pick-and-place.
Simulation makes much of this pipeline suitable for unattended execution. Human attention moves to the decisions that matter: reviewing demonstrations, checking rollouts, and comparing final success rates.
Build on It#
The registered environments are starting points. Fork the nearest task for a new embodiment or clinical procedure, then run the same record-to-validate loop. When a repeated development step has a stable command and a clear validation check, capture it as a skill. A precise description helps the coding agent route requests to it, while grounded instructions keep execution tied to the repository.
Where To Go Next#
Apply the loop to a healthcare robotics task you care about.
Add a skill for one repeated operation that still requires manual setup.
Compare policy success rates across scene randomization before moving toward hardware.
Find the full source and host requirements at https://github.com/isaac-for-healthcare/i4h-workflows under
workflows/agentic/.