Skip to content

everettVT/awesome-physical-ai

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 

Repository files navigation

awesome-physical-ai

A curated list of open source projects for Physical AI — robotics, embodied intelligence, simulation, and the infrastructure that makes it work.

Curated from the stuff I actually use, star, and interest me. PRs welcome.

Legend: ★ = in my stars · unmarked = well-known projects on my to-star list


Contents


Foundation models & policies

VLA models, behavior cloning, action models, robot foundation models.

  • Physical-Intelligence/openpi — π₀ and π₀.₅ open vision-language-action models and fine-tuning code from Physical Intelligence.
  • openvla/openvla — The open-source VLA baseline for robotic manipulation; the reference point most VLA papers compare against.
  • starVLA/starVLA — Lego-like codebase for composing and developing VLA models.
  • ginwind/VLA-JEPA — Enhancing VLA models with a latent world model (ECCV 2026).
  • NVlabs/GR00T-WholeBodyControl — Unified platform for developing and deploying humanoid whole-body control.
  • ShuangLI59/unified_video_action — Unified Video Action Model (RSS 2025): one model for video prediction and action generation.
  • x-robotics-lab/minbc — Minimal behavior cloning; the nanoGPT of imitation learning.
  • MineDojo/NitroGen — Foundation model for generalist gaming agents; embodied action models beyond robot arms.
  • huggingface/lerobot — End-to-end robot learning: datasets, policies (ACT, diffusion, SmolVLA), and real-hardware drivers.
  • NVIDIA/Isaac-GR00T — GR00T N1.x, NVIDIA's foundation model for generalist humanoid robots.
  • octo-models/octo — Transformer robot policy pretrained on 800k trajectories from Open X-Embodiment.
  • real-stanford/diffusion_policy — Visuomotor policy learning via action diffusion; spawned an entire policy-class.
  • tonyzhaozh/act — Action Chunking with Transformers, the policy behind ALOHA bimanual teleop.
  • thu-ml/RoboticsDiffusionTransformer — RDT-1B: diffusion foundation model for bimanual manipulation.
  • moojink/openvla-oft — Optimized fine-tuning recipe for VLAs; big speed and success-rate gains over vanilla OpenVLA.

Simulation & environments

Physics engines, robot sim benchmarks, multi-agent environments, digital twins.

World models

Video prediction, latent imagination, environment modeling.

  • NVIDIA/cosmos — NVIDIA's open platform of world foundation models, datasets, and tooling for Physical AI.
  • NVIDIA/Cosmos-Tokenizer — Image and video neural tokenizers that feed world-model training.
  • lucas-maes/le-wm — LeWorldModel: stable end-to-end JEPA world models from pixels.
  • danijar/dreamerv3 — Mastering diverse domains through latent imagination; the reference world-model agent.
  • facebookresearch/vjepa2 — V-JEPA 2: self-supervised video world models, Meta's bet on predictive latents over pixels.
  • eloialonso/diamond — DIAMOND: RL agents trained inside a diffusion world model.
  • 1x-technologies/1xgpt — 1X's world-modeling challenge for humanoid robots, with real robot video data.

Training & RL infrastructure

RL frameworks, GRPO, reward modeling, distributed training.

Robot operating systems & middleware

Real-time control, ROS alternatives, hardware abstraction.

  • copper-project/copper-rs — Deterministic robot OS in Rust: build, run, and replay your entire robot.
  • elodin-sys/elodin — Simulation and flight software monorepo for drones and space systems, ECS-based, in Rust.
  • mcarfagno/mpc_python — Simple MPC controller for path tracking; a readable intro to receding-horizon control.
  • IntelligentControlSystems/l4acados — Learning-based models inside acados real-time MPC.
  • bevyengine/bevy — Data-driven ECS engine in Rust; increasingly the substrate for robotics tooling and viz.
  • skypjack/entt — The C++ entity-component-system powering real-time architectures (and Minecraft).
  • meshtastic/firmware — Open LoRa mesh networking for off-grid device comms.
  • ros2/ros2 — The Robot Operating System; still the lingua franca of robot software.
  • dora-rs/dora — Dataflow-oriented robotic architecture in Rust; the ROS alternative the LeRobot crowd uses.
  • eclipse-zenoh/zenoh — Pub/sub + storage + compute protocol unifying robot-to-cloud data in motion.
  • PX4/PX4-Autopilot — The open flight-control stack for drones and VTOL.
  • ArduPilot/ardupilot — Autopilot for planes, copters, rovers, subs; two decades of field hours.
  • NVlabs/curobo — CUDA-accelerated motion planning and IK.
  • NVIDIA-ISAAC-ROS/isaac_ros_common — GPU-accelerated ROS 2 packages for NVIDIA Jetson and beyond.
  • rerun-io/rerun — Visualize, query, and stream multimodal robotics data; the log-everything debugger for embodied AI.
  • foxglove/mcap — Serialization-agnostic container format for robotics logs.
  • TheRobotStudio/SO-ARM100 — The open-hardware arm that made robot learning a $200 hobby (LeRobot's default platform).

Perception & sensing

Depth estimation, segmentation, motion capture, pose estimation.

GPU & compute primitives

Kernel programming, CUDA tooling, inference engines, quantization.

Data & datasets

Robot datasets, synthetic data, data processing for Physical AI.

Evaluation & benchmarks

VLA evaluation, sim-to-real, terminal and agent benchmarks.

Orchestration & deployment

Fleet management, cloud-to-edge, Physical AI platforms.

Papers & meta-lists


Maintained by @everettVT. If a project belongs here and isn't, open a PR.

About

A curated list of open source projects for Physical AI — robotics, embodied intelligence, simulation, and the infrastructure that makes it work.

Topics

Resources

License

Stars

1 star

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors