Developers

A few lines from robot to trained policy.

The open-source neuracore Python SDK and CLI wrap the whole workflow: log any data type, train imitation and reinforcement learning policies on the cloud, and run inference locally or on a policy server.

$pip install neuracore
record_demo.py
import neuracore as nc
import time
nc.login()
nc.connect_robot(robot_name="MyRobot", urdf_path="robot.urdf")
nc.create_dataset(name="My Robot Dataset")
# stream a demonstration into the dataset
nc.start_recording()
t = time.time()
nc.log_joint_positions(positions={"joint1": 0.5}, timestamp=t)
nc.log_rgb(name="top_camera", rgb=image, timestamp=t)
nc.log_language(name="instruction",
                language="Pick up the red cube", timestamp=t)
nc.stop_recording()  # auto-uploads to the cloud
What the SDK gives you
Logging

Streaming data logging

Log joint positions, velocities, targets, RGB cameras, language, and custom sensors with timestamps. Everything streams live to the dashboard as you record.

Robots

URDF and MuJoCo MJCF

Describe your robot once with a URDF or MJCF file, then log and train against that embodiment, single-arm, dual-arm, or dexterous.

Training

Cloud training

Launch multi-GPU runs with Diffusion Policy, ACT, or π0. Checkpoints, TensorBoard, and automatic batch-size tuning are built in.

Datasets

Access and synchronize

Load datasets, synchronize data types across embodiments at any frequency, and visualize episodes on the web dashboard.

Inference

Local and cloud policies

Load a trained policy and call predict() on your machine, on a local policy server, or on a cloud endpoint.

CLI

Command-line tooling

Authenticate, select an org, and inspect or monitor cloud and local runs. Launch a policy server to sanity-check inference.

Train and run

Kick off cloud training, then call the policy.

Launch a multi-GPU run with Diffusion Policy, ACT, or π0 against a dataset, then load the trained policy and call predict() locally, on a policy server, or on a cloud endpoint.

Imitation vs. reinforcement learning
train_and_infer.py
# train on the cloud
nc.start_training_run(
    name="MyTrainingJob",
    dataset_name="My Robot Dataset",
    algorithm_name="diffusion_policy",
    num_gpus=5,
    frequency=50,
)
# load the trained policy and predict
policy = nc.policy(train_run_name="MyTrainingJob")
nc.log_joint_positions(positions={"joint1": 0.5})
nc.log_rgb(name="top_camera", rgb=image)
actions = policy.predict(timeout=5)
terminal
neuracore login
neuracore select-org --org-name "Acme Robotics"
neuracore training list --limit 5
neuracore training monitor --training-name my_experiment
Command line

Monitor runs from the terminal.

Authenticate, pick your organization, and inspect or monitor cloud and local training runs. TensorBoard opens straight from the CLI.

Get Started

Install the SDK and log your first demonstration.

Read the docs, clone a repo, or run the getting-started notebook on Colab.