Your First RL Training Job

Submit an Isaac Lab RL training job to OSMO and verify that training metrics appear in MLflow. By the end of this recipe, you will have a trained checkpoint for the Anymal-C velocity locomotion task.

[!NOTE] This recipe requires deployed infrastructure with OSMO running. Complete the Quickstart first.

📋 Prerequisites

Requirement	Details
Infrastructure	Azure resources deployed via Terraform
OSMO	Control plane and backend running (kubectl get pods -n osmo-control-plane)
VPN	Connected to private cluster (if using private AKS)
Azure CLI	Authenticated (az login)
kubectl	Connected to AKS cluster

🚀 Steps

Step 1: Preview the training configuration

The submission script auto-detects Azure context from Terraform outputs. Preview what will be submitted before running a real job:

cd training/rl/scripts
./submit-osmo-training.sh --config-preview

Review the output. The preview shows the workflow template, task name, container image, GPU/CPU/memory allocation, and Azure ML context. No job is submitted.

Step 2: Run a smoke test (optional)

Verify Azure connectivity and credential resolution before committing GPU time:

./submit-osmo-training.sh \
  --task Isaac-Velocity-Rough-Anymal-C-v0 \
  --num-envs 16 \
  --max-iterations 10 \
  --run-smoke-test

The smoke test validates Azure ML workspace access, MLflow tracking URI resolution, and storage connectivity. It runs a minimal training loop (10 iterations, 16 environments) to catch configuration issues early.

Step 3: Submit a training job

Submit a full training run with the SKRL backend (default):

./submit-osmo-training.sh \
  --task Isaac-Velocity-Rough-Anymal-C-v0 \
  --num-envs 2048 \
  --max-iterations 1500

The script packages training/rl/ as a base64 payload, injects it into the OSMO workflow template, and submits the job. OSMO schedules the job on a GPU node via KAI Scheduler.

Override the backend to use RSL-RL instead of SKRL:

./submit-osmo-training.sh \
  --task Isaac-Velocity-Rough-Anymal-C-v0 \
  --backend rsl_rl \
  --num-envs 2048 \
  --max-iterations 1500

Step 4: Monitor training progress

OSMO UI: Open the OSMO dashboard to view workflow status, pod logs, and real-time metrics. See Accessing OSMO for connection instructions (VPN or port-forward).

Azure ML Studio: Navigate to your workspace at ml.azure.com, open the Jobs section, and select the MLflow experiment. Training metrics (reward, episode length, loss) stream in real time as the job progresses.

CLI (optional): Check workflow status via the OSMO CLI:

osmo workflow list

Step 5: Verify results in MLflow

1. Open ml.azure.com and select your workspace
2. Navigate to Jobs in the left sidebar
3. Find your experiment (named after the task, e.g., isaaclab-training) and select it
4. Click the latest run to open the run detail page
5. Select the Metrics tab to view training curves (reward, episode length, loss) — use the chart controls to overlay multiple metrics or smooth noisy curves
6. Select the Outputs + logs tab to view stdout/stderr logs from the training container
7. If --register-checkpoint was used, navigate to Models in the left sidebar to confirm the registered model and its version

List registered models from the CLI:

az ml model list \
  --resource-group <your-resource-group> \
  --workspace-name <your-workspace>

✅ Verify

The recipe succeeded when:

• OSMO pod reached Completed status
• MLflow experiment shows training metrics across iterations
• A checkpoint artifact exists in the Azure ML model registry (if --register-checkpoint was used)

⚙️ Configuration Reference

Parameter	Default	Description
--task	Isaac-Velocity-Rough-Anymal-C-v0	Isaac Lab task environment
--num-envs	2048	Parallel simulation environments
--max-iterations	(unset)	Training iterations; omit for task default
--backend	skrl	Training backend (skrl or rsl_rl)
--gpu	1	GPU count
--checkpoint-mode	from-scratch	from-scratch, warm-start, or resume
--register-checkpoint	(none)	Model name for Azure ML registration

See Scripts Reference for the full parameter table.

🔗 Related Recipes

• Your First LeRobot Training Job — behavioral cloning alternative
• End-to-End LeRobot Pipeline — automated train → evaluate → register
• Isaac Lab Training Guide — detailed RL reference documentation

🤖 Crafted with precision by ✨Copilot following brilliant human instruction, then carefully refined by our team of discerning human reviewers.