Deployment in Simulation#

Serve the trained policy from a headless daemon, run Arena episodes, and record rollouts for human review.

Learning Objectives#

By the end of this lesson, you’ll be able to:

Evaluate any shipped environment or your own checkpoint from a single prompt.
Explain how the policy daemon and Arena work together to run a rollout.
Record verification episodes for human review.

Run It With a Prompt#

Evaluate the checkpoint from the previous lesson with the same stack used for training. N1.5 is the shipped route. The N1.7 option explicitly confirms or restores the newer stack and model before launch.

N1.5: Evaluate (Shipped Default)

Evaluate scissor_pick_and_place for 2 episodes using the checkpoint I just trained.

N1.7: Evaluate (Optional Upgrade)

Confirm scissor_pick_and_place uses policy.stack gr00t_n17 and policy.model_repo
nvidia/SO_ARM_Starter_Gr00tN17, updating the YAML if needed. Evaluate it for 2 episodes
using the GR00T N1.7 checkpoint I just trained. Verify the N1.7 route and checkpoint path
before launch, record the rollouts, and report the task success rate.

When review is complete:

Stop all

What Happens Under the Hood?#

Policy evaluation runs two processes: a policy daemon and Arena. Rollouts can be recorded to HDF5. State-machine smoke tests use Arena’s controller without a policy daemon.

How Deployment Fits Together#

The policy daemon loads the foundation-model stack and serves inference without a window.
Arena is the simulation. It steps the environment, queries the policy, and is the only process that opens the sim window.

The top-level policy/run.sh routes to the right stack by env id (policy.stack in the env YAML):

Subproject	Environments
GR00T N1.5 (`gr00t_n15/`)	Shipped default for `scissor_pick_and_place`, plus `assemble_trocar` and `surgical_*` policy baselines
GR00T N1.7 (`gr00t_n17/`)	Optional `scissor_pick_and_place` upgrade selected by editing the env YAML
GR00T N1.6 (`gr00t_n16/`)	`locomanip_tray_pick_and_place`, `locomanip_push_cart`
openpi PI0 (`openpi_pi0/`)	`ultrasound_liver_scan`

The N1.5 path uses the repository configuration as shipped. The N1.7 path updates the scissor environment’s policy.stack and policy.model_repo, then passes the matching checkpoint to the N1.7 inference service.

For policy evaluation, launch order matters: start the daemon first, wait until it logs policy ready, then launch Arena. If Arena starts first, the episode can’t connect. For state-machine smoke runs, Arena owns the controller and records the episode directly.

Optional deep dive:

Explain how the policy daemon and Arena communicate over Zenoh during a rollout, and why the daemon must report ready before Arena starts.

Running a Rollout#

The eval prompt handles the orchestration: the daemon starts first, and once it logs policy ready, Arena runs the episodes against it and records each one for review. To evaluate your own fine-tuned checkpoint instead of the shipped model, the prompt simply points the daemon at your checkpoint directory.

Optional deep dive:

Walk me through what happens each control step when Arena queries the live policy and applies the returned action chunk.

What’s Next?#

You can now deploy and validate a policy entirely in simulation. Next, you’ll step back and assemble these pieces into a workflow of your own in Build Your Own Workflow.