RoboActions

1. Initialize RoboActions

Connect to the RoboActions SDK with your API key. Your gateway to powerful AI for robotics.

import roboactions

# Initialize the RoboActions SDK
sdk = roboactions.init(
  api_key="YOUR_API_KEY",
  robot_id="YOUR_ROBOT_ID" # Link to your registered robot
)

2. Easy Data Collection

Define sensor configurations and seamlessly collect data in your robot's loop. Upload to RoboActions Cloud.

# Initialize a Dataset object.
dataset = sdk.new_dataset(
  name="my_robot_data",
  sensor_config={ # Define your robot's sensors
    "camera_main": {"resolution": (640, 480), "fps": 30, "type": "RGB"},
    "wrist_camera": {"resolution": (320, 240), "fps": 15, "type": "Depth"},
    "lidar_2d": {"enabled": True, "range_m": 10},
    "joint_states": {"num_joints": 6},
    # Add more sensors as per your robot's configuration
  }
)

# Collect data in your robot's control loop.
for step in range(100): # Example loop
  observations = robot.get_current_observations() # Your function to get sensor data
  actions = robot.get_last_actions()           # Your function to get actions taken
  
  # Record observation-action pair at each timestep
  dataset.record(timestamp=time.time(), observations=observations, actions=actions)
  time.sleep(0.1) # Match your robot's control frequency

# Save and upload your dataset to RoboActions Cloud
# SDK handles chunking, retries, and data integrity.
dataset_id = dataset.save(upload=True)
print(f"Dataset '{dataset.name}' uploaded successfully. ID: {dataset_id}")

3. Scale Up with MimicGen

Amplify your dataset using MimicGen to improve model robustness and real-world performance.

# Augment data with MimicGen for greater diversity
augmented_dataset_id = sdk.mimicgen(
  dataset_id=dataset_id,
  num_variants=500,       # Generate 500 augmented trajectories
  augmentation_options={
    "noise_level": 0.02,   # Add slight sensor noise
    "domain_shift": True,  # Simulate different environments
    "temporal_jitter": 0.1 # Vary action timing slightly
  }
)
print(f"Augmented Dataset ID: {augmented_dataset_id}")

4. Fine-tune and Deploy

Fine-tune leading Vision-Language-Action (VLA) models like OpenVLA or pi0 on your data. Deploy for inference with ease.

# Fine-tune a foundation model (e.g., OpenVLA, pi0)
model = sdk.finetune(
  base_model="OpenVLA", # or "pi0", "N1-Gr00t"
  dataset_id=dataset_id,
  augmented_dataset_id=augmented_dataset_id, # Optional
  epochs=10,
  batch_size=32,
  learning_rate=1e-4,
  # gpu="nvidia-a100",    # Specify GPU or use 'auto'
  deploy_on_completion=True # Automatically deploy for inference
)
print(f"Model '{model.name}' fine-tuned. Accuracy: {model.accuracy*100:.2f}%")
print(f"Deployed Model ID: {model.model_id}")

5. Hybrid Inference

Run your models on-robot for low latency or in the cloud for heavy tasks. Optimal performance, your choice.

# Run hybrid inference (on-robot if capable, otherwise cloud)
current_obs = robot.get_current_observations()
task_prompt = "Pick up the red block and place it in the green bowl."

prediction = sdk.predict(
  model_id=model.model_id,
  observations=current_obs,
  task_prompt=task_prompt,  # For VLAs
  mode="hybrid"             # 'hybrid', 'cloud', or 'edge'
)

# prediction.action_sequence contains predicted future actions
next_action = prediction.action_sequence[0]
robot.execute_action(next_action)