Software stack for Physical AI

The open-source ecosystem for Physical AI is growing rapidly, transitioning from legacy robotics frameworks (like raw ROS packages) into unified machine learning stacks.

The most prominent GitHub repositories organizing and building the Physical AI and Robotics Android Stack are grouped below by layer.

📚 1. Curated Lists & Architecture Overviews (The Framework Aggregators)

These repositories provide highly detailed engineering maps.

natnew / awesome-physical-ai: A comprehensive, engineering-oriented roadmap tracking the full Physical AI stack. It maps everything from simulation engines, generalist policies, data collection patterns, to production-grade deployment. [2]
keon / awesome-physical-ai: Focuses closely on the academic and research side, cataloging essential papers, world models, and robotic foundation models. [3, 4]
DravenALG / awesome-vla-wam: Dedicated entirely to tracking Vision-Language-Action (VLA) models and World Action Models (WAM). [5]

🧠 2. Vision-Language-Action (VLA) & System 2 Layer

These repositories contain actual open-source weights, tokenizers, and training scripts for the cognitive "brain" of robots. [6, 7]

Physical-Intelligence / openpi: Published by the Physical Intelligence team, this repo holds the open-source π₀ (Pi-Zero) model family. It features a flow-based vision-language-action model (π₀) and an autoregressive version (π₀-FAST) used to drive multiple different robot bodies using unified action tokenizers.
openvla / openvla: An open-source, 7-billion-parameter VLA model designed specifically for robotic manipulation. It provides fine-tuning scripts and PyTorch implementations to map camera streams directly to discrete motor instructions.
OpenHelix-Team / OpenHelix: An open-source, dual-system VLA model explicitly aimed at real-world deployment on single-arm, dual-arm, and humanoid robotics.
OpenDriveLab / WholebodyVLA: Codebase for unified latent VLA models managing whole-body loco-manipulation. It handles coordinating walking patterns while simultaneously performing precision hand movements. [6, 7, 8, 9, 10]

⚡ 3. Physical Grounding & Simulation (System 1 Layer)

To train robots safely before applying them to physical hardware, massive physics-grounded simulators are used. [11, 12]

isaac-sim / IsaacLab: The central, open-source framework by NVIDIA for robotics research. It is highly optimized for GPU acceleration, Reinforcement Learning (RL), and fine-tuning "sim-to-real" pipelines. [13]
Lab-of-AI-and-Robotics / IsaacLab-Tutorial: A highly regarded 10-chapter developer community project detailing how to write reinforcement learning code in Isaac Lab, scaling up from simple quadrupeds to complex humanoid locomotion. [14]

⚙️ 4. Middlewares & Automation Toolchains (The Abstraction Layer)

These projects act as the connective tissue—or "Android operating system"—translating foundation models into safe hardware runtime execution. [15]

ros-physical-ai / SIG Repository: Hosted under the Open Source Robotics Alliance (OSRA) ecosystem. This Special Interest Group is standardizing declarative interfaces for ROS 2 so that arbitrary foundation AI models can easily control physical actuators.
RobotecAI / rai: An open-source, vendor-agnostic agentic framework for Physical AI. It leverages ROS 2 tools to allow LLM agents to execute real-world multi-step physical scenarios, parse system logs, and interact via voice.
microsoft / physical-ai-toolchain: A production-grade framework integrating cloud services with edge environments. It streamlines the heavy lifting of physical data curation, data augmentation, and model edge deployment directly onto NVIDIA Jetson hardware. [1, 16, 17]

Reference:

V-Modal AI

DEV Community