How neuracoreai/neuracore works
Project type and purpose
• This is an open-source Python client library/SDK (“Neuracore Python Client”) plus training infrastructure for robotics data collection, streaming, and ML model development/deployment.
• It provides:
• A CLI layer (commands like nc-login, nc-launch-server, neuracore-validate)
• An API client for interacting with the Neuracore SaaS platform (HTTP/REST, SSE, P2P streaming)
• Core abstractions for robots, datasets, recording, streaming, and CLI utilities
• A modular ML training subsystem (Hydra configurations, PyTorch DDP, built-in algorithms, custom-algorithm plugin API)
File structure
• Top-level directories:
• neuracore/api: HTTP client, endpoints, datasets, training, logging
• neuracore/core: authentication, config manager, CLI commands, data/streaming abstractions, utilities
• neuracore/ml: training entrypoint, config/, algorithms/, datasets/, trainers/, utils/
• dataset_upload_scripts: standalone scripts for bulk dataset upload
• examples/: code samples (Python, ROS/Docker) demonstrating usage
• tests/: unit & integration tests
• Config & build: setup.py, cSpell.json, .github CI/deploy workflows, pre-commit, CHANGELOG, LICENSE
• Patterns:
• Layered architecture (CLI → core → api)
• Plugin pattern for algorithms (dynamic discovery via algorithm_utils)
• Hydra for config composition in training
• Streaming pipelines (SSE, P2P) modularized under core/streaming
Key insights from README
• High-level flow: user authenticates → connects to robot → logs data (joint states, images, language) → streams to SaaS or local endpoint → trains models (local or cloud) → deploys inference endpoints
• Supported technologies: URDF, MuJoCo, PyTorch, Hydra, SSE, custom data types
• CLI tools provide entrypoints into core functionality