End-to-End LeRobot Pipeline

Run the full LeRobot pipeline — train a policy, evaluate it against simulation episodes, and register the model to Azure ML — in a single command. The pipeline script handles workflow submission, status polling, and stage transitions automatically.

[!NOTE] Complete Your First LeRobot Training Job before this recipe to verify that single-stage submission works.

📋 Prerequisites

Requirement	Details
Infrastructure	Azure resources deployed via Terraform
OSMO	Control plane and backend running
Basic LeRobot recipe	Single-stage training verified successfully
HuggingFace account	Write access to a policy repo for pushing trained weights

🚀 Steps

Step 1: Understand the pipeline stages

The run-lerobot-pipeline.sh script orchestrates three stages:

Train → Wait → Evaluate → Register
  │              │           │
  │              │           └── Register model to Azure ML
  │              └── Submit inference/eval workflow
  └── Submit training workflow, poll until complete

Each stage submits an OSMO workflow and polls for completion before advancing.

Step 2: Preview the pipeline configuration

cd training/pipelines
./run-lerobot-pipeline.sh \
  -d lerobot/aloha_sim_insertion_human \
  --policy-repo-id <your-hf-username>/aloha-act-policy \
  --config-preview

The preview shows both training and inference configurations, polling intervals, and timeout settings.

Step 3: Run the full pipeline

./run-lerobot-pipeline.sh \
  -d lerobot/aloha_sim_insertion_human \
  --policy-repo-id <your-hf-username>/aloha-act-policy \
  -r my-aloha-act-model

This command:

1. Submits an ACT training job with the ALOHA sim insertion dataset
2. Polls OSMO every 60 seconds until training completes (default timeout: 720 minutes)
3. Submits an evaluation workflow against the trained policy
4. Registers the model as my-aloha-act-model in Azure ML

Step 4: Customize pipeline parameters

Adjust training and evaluation settings:

./run-lerobot-pipeline.sh \
  -d lerobot/aloha_sim_insertion_human \
  --policy-repo-id <your-hf-username>/aloha-act-policy \
  --policy-type act \
  --training-steps 50000 \
  --save-freq 5000 \
  --eval-episodes 20 \
  --poll-interval 120 \
  --timeout 360 \
  -r my-aloha-act-model

Step 5: Run training only (skip evaluation)

Use --skip-inference when iterating on training hyperparameters:

./run-lerobot-pipeline.sh \
  -d lerobot/aloha_sim_insertion_human \
  --policy-repo-id <your-hf-username>/aloha-act-policy \
  --skip-inference

Step 6: Run in async mode

Submit training without waiting for completion:

./run-lerobot-pipeline.sh \
  -d lerobot/aloha_sim_insertion_human \
  --policy-repo-id <your-hf-username>/aloha-act-policy \
  --skip-wait

Check status manually through the OSMO UI or pod logs.

✅ Verify

The recipe succeeded when:

• Training pod completed successfully
• Evaluation pod completed with success-rate metrics logged to MLflow
• Model appears in Azure ML registry:

az ml model show \
  --name my-aloha-act-model \
  --resource-group <your-resource-group> \
  --workspace-name <your-workspace>

⚙️ Configuration Reference

Parameter	Default	Description
-d, --dataset-repo-id	(required)	HuggingFace dataset repository
--policy-repo-id	(required)	HuggingFace repo for trained policy
--policy-type	act	Policy architecture (act or diffusion)
--training-steps	(task default)	Total training iterations
--eval-episodes	10	Evaluation episodes
--poll-interval	60	Status check interval in seconds
--timeout	720	Training timeout in minutes
--skip-inference	(disabled)	Skip evaluation stage
--skip-wait	(disabled)	Async mode — submit without waiting
-r, --register-model	(none)	Model name for Azure ML registration

See Scripts Reference for the full parameter table.

🔗 Related Recipes

• Your First LeRobot Training Job — single-stage training
• Your First RL Training Job — reinforcement learning alternative
• Preparing Datasets for Training — dataset download and validation

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