Open-source repositories for robotics and physical AI teams, from the robot software platform to safety and reliability.
Pick your assistant, copy the prompt and paste it into Claude, ChatGPT, Gemini or Grok.
Claude
Read https://automotivecloudwatch.com/repositories/robotics.md and map it to my team. Ask me up to four short questions first: my team, my program and its stage, what we already run, and any hard constraints. Then give me one written assessment: a short summary, three maps (already running, not running yet, still done by hand), the top 10 repositories rated for impact and effort, a setup check, and a 30-day plan that starts with the one repository to try today.ChatGPT
Open https://automotivecloudwatch.com/repositories/robotics.md with your browsing tool and read all of it, then map it to my team. Ask me up to four short questions first: my team, my program and its stage, what we already run, and any hard constraints. Then give me one written assessment: a short summary, three maps (already running, not running yet, still done by hand), the top 10 repositories rated for impact and effort, a setup check, and a 30-day plan that starts with the one repository to try today. If browsing is off or the link will not open, tell me and I will paste the file.Gemini
Read https://automotivecloudwatch.com/repositories/robotics.md and map it to my team. If you cannot open the link, tell me and I will paste the file. Ask me up to four short questions first: my team, my program and its stage, what we already run, and any hard constraints. Then give me one written assessment: a short summary, three maps (already running, not running yet, still done by hand), the top 10 repositories rated for impact and effort, a setup check, and a 30-day plan that starts with the one repository to try today.Grok
Fetch and read https://automotivecloudwatch.com/repositories/robotics.md, then map it to my team. If you cannot open the link, say so and I will paste the file. Ask me up to four short questions first: my team, my program and its stage, what we already run, and any hard constraints. Then give me one written assessment: a short summary, three maps (already running, not running yet, still done by hand), the top 10 repositories rated for impact and effort, a setup check, and a 30-day plan that starts with the one repository to try today.You get a written assessment of your team: a short summary, three maps of what you run, what you lack and what you still do by hand, the top 10 repositories rated for impact and effort, a setup check and a 30-day plan.
Your assistant will also offer, with one yes or no question, to keep watch: a new edition lands every two weeks, and on a yes it reads each one the day after and tells you only what changed for your team. You can cancel anytime.
Plain text for an assistant: /repositories/robotics.md /repositories/robotics.json
Edition 1, checked 2026-10-19
A robot demo can now be assembled almost entirely from public code. A robot that works a full shift cannot, and the gap between the two is a curation problem.
The public stack has grown fast. Robot foundation models, open teleoperation datasets, GPU simulators and humanoid training frameworks now sit beside ROS 2 and the classic planners. Stars tell you which of them are getting attention. They do not tell a fleet team which navigation stack will hold up in a crowded warehouse, or tell a learning team whose dataset loader will still be maintained after the paper that launched it is two years old.
Edition 1, checked 2026-10-19, lists 42 repositories across 8 teams. By level that is 11 Start here, 20 Practitioner, 11 Deep end.
Robot Learning and Foundation Model has the most entries, 7; Perception; Robot Safety and Reliability have the fewest, 4. 6 entries carry a license flag (weak or strong copyleft, a non-standard license or no license detected), so read the license before you plan around them. Last activity runs from 2024-12-24 to 2026-10-19, as checked on 2026-10-19. 6 entries have had no push in over 12 months (droid-dataset/droid, Lifelong-Robot-Learning/LIBERO, real-stanford/diffusion_policy, openvla/openvla, dexsuite/dex-retargeting, unitreerobotics/unitree_rl_gym); each shows its last activity date.
This is the first edition, so nothing has moved yet.
Robot Software Platform 5 Motion Planning and Control 5 Robot Learning and Foundation Model 7 Simulation and Sim-to-Real 6 Manipulation and Humanoid 6 Mobile Robot and Fleet 5 Perception 4 Robot Safety and Reliability 4
Own the ROS 2 distribution, middleware choices, hardware abstraction and logging formats that the rest of the robot software runs on.
5 repositories
The entry repository for ROS 2 that lists the core packages and points to the documentation and release process.
Start here to get the full picture of the ROS 2 distribution, and note that the license lives in the individual packages, which are mostly Apache-2.0.
A framework for connecting ROS 2 controllers to real or simulated hardware through a common hardware interface.
Use it when you integrate a new robot or actuator so that standard controllers, MoveIt and Nav2 work without custom glue.
A ROS 2 middleware layer that uses Zenoh instead of DDS for discovery and transport.
Try it when DDS discovery struggles on Wi-Fi, across networks or with large fleets, which are common pain points in deployed robots.
An open container file format for timestamped multimodal robot data, with libraries in several languages.
It is now the default rosbag2 storage format and a good choice for any robot or vehicle logging pipeline that needs fast indexed reads.
A dataflow framework written in Rust for composing robot and AI nodes with shared-memory messaging, usable from Python, Rust and C++.
Look at it when ROS 2 overhead or Python integration of AI models is the bottleneck and you can accept a younger ecosystem.
Own kinematics, dynamics, trajectory optimization and collision-aware planning that turn tasks into safe, feasible motion.
5 repositories
The standard ROS 2 framework for manipulator motion planning, collision checking and grasp execution.
It is the usual starting point for any arm application on ROS 2 and connects to most other planning tools on this list.
A library of sampling-based motion planning algorithms such as RRT and PRM variants.
Learn it to understand what MoveIt does underneath and to try planners directly on new robots or state spaces.
A fast C++ and Python library for rigid body kinematics, dynamics and their analytical derivatives.
Reach for it when building model predictive control or trajectory optimization, where fast dynamics derivatives decide performance.
A GPU-accelerated library for collision-free motion generation, inverse kinematics and trajectory optimization.
Use it when planning time matters, for example in high-throughput picking cells, since it runs many optimizations in parallel on a GPU.
A toolbox for model-based design and verification of robots, combining multibody dynamics, optimization and systems modeling.
It is the tool to study when you need rigorous dynamics, contact modeling and mathematical programming in one framework.
Own the data pipelines, training code, benchmarks and pretrained vision-language-action models used to teach robots new skills.
7 repositories
Hugging Face's library for real-world robot learning, covering data collection, a shared dataset format, training and deployment of policies.
It is the easiest end-to-end entry point into imitation learning and VLA fine-tuning on affordable hardware.
The home of the Open X-Embodiment dataset, which pools robot demonstrations from many labs and robot types, along with RT-1-X model code.
Read it to understand the cross-embodiment data that most current open VLA models were pretrained on.
Quiet for over 12 months
The code for the DROID platform and its large in-the-wild Franka manipulation dataset collected across many sites.
Study it for its data collection hardware and software setup, which is a strong template for building your own demonstration pipeline.
Quiet for over 12 months
A simulated benchmark of language-conditioned manipulation task suites for lifelong and multitask robot learning.
Many VLA papers report results on it, so it is a practical place to reproduce claims and compare models on equal terms.
Quiet for over 12 months
The reference code for Diffusion Policy, which generates robot actions with a denoising diffusion model trained on demonstrations.
It is a canonical reference that many later VLA action heads build on, although the code itself has not been updated since December 2024.
Quiet for over 12 months
An open 7B-parameter vision-language-action model with training and fine-tuning code.
Use it to learn how a VLA is assembled from a vision-language model and action tokens, and how to fine-tune one on your own data.
Physical Intelligence's open release of its pi-zero family of VLA models with weights, fine-tuning and inference code.
It is one of the strongest open generalist policies to fine-tune for a new robot or task.
Own physics engines, robot models and GPU training environments, and the process of moving policies from simulation onto hardware.
6 repositories
A fast and accurate physics engine for robotics and biomechanics, maintained by Google DeepMind.
It is the common physics baseline for robot learning research and runs on CPU or, through MJX and Warp, on accelerators.
A curated collection of tuned MuJoCo models of commercial robots, arms, hands and humanoids.
Use it to avoid weeks of model building and tuning, and check each model folder for its own license.
NVIDIA's open framework on top of Isaac Sim for GPU-parallel robot learning, including reinforcement and imitation learning.
It is a common path for training locomotion and manipulation policies at scale and transferring them to hardware.
A GPU physics engine built on NVIDIA Warp and hosted by the Linux Foundation, with contributions from Disney Research, Google DeepMind and NVIDIA.
Watch it as the likely shared physics core for future Isaac Lab and MuJoCo workflows, with differentiable simulation built in.
A Python physics platform for robotics and embodied AI that combines several solvers with fast parallel simulation and rendering.
Evaluate it as an alternative training environment, and benchmark it on your own tasks before relying on headline speed claims.
The core library of Gazebo, the long-standing open robot simulator tightly integrated with ROS 2.
Use it for system-level testing of ROS 2 robots with sensors and worlds, where learning-focused simulators are less complete.
Own arms, hands and humanoid platforms, including inverse kinematics, retargeting from human motion and whole-body control.
6 repositories
Open hardware designs for low-cost SO-100 and SO-101 robot arms that work directly with LeRobot.
It gives teams a cheap physical platform to try imitation learning before spending on industrial arms.
A Python library for differential inverse kinematics on MuJoCo models with tasks, limits and priorities.
Reach for it when you need clean, constraint-aware IK for arms, hands or humanoids in prototypes and teleoperation.
Quiet for over 12 months
Tools for mapping human hand poses from video or motion capture onto different robot hands.
It is a key piece for teleoperation and for turning human demonstrations into training data for dexterous hands.
Quiet for over 12 months
Unitree's reinforcement learning examples for training and deploying locomotion policies on its Go2, H1 and G1 robots.
It shows the full sim-to-real loop for legged and humanoid locomotion on hardware many labs actually own.
Research code from CMU for aligning humanoid simulation with real physics so that agile whole-body skills transfer to hardware.
Read it to understand current methods for closing the sim-to-real gap on humanoids beyond plain domain randomization.
NVIDIA's open GR00T N foundation models for humanoid and general robot control, with fine-tuning and deployment code.
Look at it when targeting humanoids or the NVIDIA Jetson Thor stack, since it ties into Isaac Lab and NVIDIA data tools.
Own AMR navigation, mapping for navigation and the fleet control layer, including VDA 5050 integration and multi-vendor traffic management.
5 repositories
Nav2, the ROS 2 navigation framework for planning, control, recovery behaviors and behavior trees on mobile robots.
It is the standard open base for AMR navigation and is used in many commercial robots, with mixed Apache-2.0, BSD and LGPL licensing by package.
A 2D laser SLAM package for ROS 2 that supports mapping, localization and multi-session map merging.
It is the default mapping tool used with Nav2 for warehouse and indoor mobile robots.
The official source of the VDA 5050 interface specification for communication between AGVs or AMRs and a master control system, now at version 3.0.0.
Read it before any multi-vendor fleet project, since it defines the MQTT messages and state model that vendors are expected to implement.
A Fraunhofer IML open fleet control system for AGVs and AMRs with built-in VDA 5050 support.
Use it to prototype or run vendor-neutral fleet control, including routing, order management and traffic handling.
The ROS 2 core of Open-RMF, a framework for coordinating fleets of different robots with shared infrastructure such as doors and lifts.
It fits facilities that must run robots from several vendors together and manage traffic and task allocation between them.
Own state estimation, odometry, SLAM and 3D data processing that tell a robot where it is and what is around it.
4 repositories
A C++ library with Python bindings for smoothing and mapping with factor graphs.
It is the estimation back end behind many SLAM and sensor fusion systems, so learning it pays off across localization work.
A simple LiDAR odometry pipeline based on point-to-point ICP that works across sensors with little tuning.
Use it as a strong baseline before reaching for more complex LiDAR SLAM systems.
A GPU-accelerated stereo visual-inertial odometry and SLAM package for ROS 2 on NVIDIA Jetson.
Use it when you need camera-based localization at frame rate on embedded NVIDIA hardware.
A C++ and Python library for 3D data processing, including point cloud registration, surface reconstruction and visualization.
Keep it in the toolbox for inspecting, filtering and aligning point clouds while developing LiDAR and depth camera pipelines.
Own security hardening, real-time tracing, safe-control benchmarking and vulnerability awareness across the robot software stack.
4 repositories
Tools for enabling and managing DDS security in ROS 2, including keystores, certificates and access control policies.
Use it to add authentication, encryption and access control to ROS 2 traffic, which is required for any robot on a shared network.
Instrumentation and tools for low-overhead tracing of ROS 2 systems with LTTng.
Reach for it when you need evidence of end-to-end latency and executor behavior for real-time or safety analysis.
A benchmark suite from the University of Toronto for comparing control and reinforcement learning methods under safety constraints.
Use it to evaluate how learned controllers behave against constraints and disturbances before trusting them on hardware.
The Robot Vulnerability Database, a public record of security flaws and bugs found in robots and robot software.
Check it during threat modeling and supplier reviews to see the kinds of weaknesses that have already been found in common platforms.
Suggest a repository or a correction
Want a note when a new edition lands? Subscribe free on the ACW Substack. Subscribe free
The other libraries: Automotive Manufacturing All repositories