01Open source
Repos
Open-source repos worth knowing, by the problem they take on.
Robotics
Mobile Manipulator Stack
ROS 2 and Isaac Sim reference stack for a low-cost mobile manipulator with dual SO-101 arms, RGB-D camera, 2D LiDAR, and IMU. It integrates modeling, simulation, control, planning, perception, data QC, and imitation learning for practical hardware use.
Nav2 Tuning GUI
A ROS 2 GUI for tuning Nav2 navigation parameters on a live stack without restarting nodes; operators can adjust values such as max_vel_x and inflation_radius and see effects immediately.
Quackd Robot CLI
A CLI for connecting and commanding robots on LeRobot, AlohaMini, ToddlerBot, or ROS bases with LLM brains from Claude, OpenAI, Gemini, Grok, or local models; includes simulator, .duck safety contracts, MCP, memory, and flock coordination.
KERV VLA Models
Official code for the DAC 2026 paper 'KERV: Kinematic-Rectified Speculative Decoding for Embodied VLA Models.' It accelerates autoregressive vision-language-action models via speculative decoding that connects token-domain generation with kinematic-domain constraints.
FastCrest Tether CLI
FastCrest Tether is an open-source edge-to-cloud CLI for optimizing, verifying, and deploying VLA models across Jetson, RTX, Apple Silicon, and AMD. It certifies parity of ONNX exports against reference PyTorch policies (cosine +1.000000) on SmolVLA, pi0, pi0.5, and GR00T N1.6.
EgoExoMoCap Motion Capture
Official code for the ECCV 2026 paper EgoExoMoCap, a distributed ego-exo human motion capture method. It is notable as research code relevant to humanoid teleoperation.
LeHome Challenge Solution
Solution for the LeHome Challenge 2026 on bimanual garment folding with a LeRobot SO-ARM101, placing 1st of 62 online and 2nd in the real-world final. It is notable as a top competition entry for home-robot manipulation.
OpenAMR Platform Software
ROS 2 Jazzy software stack for the OpenAMRobot mobile base, including simulation, Nav2, SLAM, docking, control, drivers, and bringup. It is notable as an actively maintained full mobile-base stack with AprilTag-based autodocking.
RIMKit Motion Retargeting
RIMKit converts SOMA human motion into whole-body motion for humanoid robots, supporting sixteen robot targets from Unitree, ROBOTIS, Apptronik, and others via Python and CLI. It is notable as a practical contact-aware retargeting toolkit for humanoid developers.
PhysInOne Python
Python code for PhysInOne, a CVPR 2026 dataset of 2 million videos from 153,810 dynamic 3D scenes covering 71 physical phenomena. It is notable as the largest physically-grounded training dataset for AI systems.
Unitree G1 Autonomy
ROS 2 Jazzy autonomy stack for the Unitree G1 humanoid, developed simulation-first against unitree_mujoco. It is notable because the simulator uses the same DDS channels as the hardware, letting navigation and control code transfer unchanged.
Visloc Rust SLAM
Pure Rust visual and visual-inertial SLAM, SfM, and localization library. It is notable for benchmarking faster than COLMAP on ETH3D scenes while achieving lower camera-centre RMSE.
Robotics Skills Suite
Robotics Skills Suite provides 38 builder and 38 reviewer Claude skills that automate xlsx deliverables for industrial-robot lifecycles. It pairs each builder skill with an audit-ready reviewer, covering ISO 10218, 13849, 62061, 12100, and IEC 62443.
CHEK EGO Miner
CHEK EGO Miner is a Qingkong Technology workflow using the CHEK mobile app and Qingkong Miker desktop app to collect first-person EGO data for embodied AI. It lowers robotics data collection to everyday users recording real work.
Autonomous OS
Autonomous OS is an open-source robot operating system that runs an on-robot agentic reasoning engine to decide actions from sensor input. It is notable for making robots autonomous rather than remote-controlled or scripted, with skills like guarding and face recognition.
Spectacles Dimensional OS
Spectacles-dimensional-os integrates Snap Spectacles AR glasses with Dimensional OS to control Unitree GO2 or G1 robots for navigation and sensor visualization. It is notable for exploring augmented reality as an interface for physical AI and robotics tasks.
Watti Desk Lamp
Watti is a prototype desk lamp that plays light scenes and is being extended with depth sensing, movement, scanning, and desktop-assistant features. The public repository documents the project while its code is not yet available.
Phyzical AI Data
Phyzical_org is a browser-based teleoperation platform that lets users control simulated robot arms to produce 30–60 Hz demonstration trajectories for embodied AI. It is notable for generating elizaOS-ready episodes with on-chain provenance from casual browser users.
SimART Platform
SimART is an open-source platform for wireless communication and sensing research built on ROS1, combining Sionna ray tracing, AirSim/Unreal visualization, and rosbag replay. It is notable for supporting digital-twin scenes, trajectory replay, and beam selection evaluation in 3D environments.
CoC Autolabeler
NVIDIA CoC Autolabeler is an open pipeline that generates Chain-of-Causation reasoning labels and meta-actions from raw autonomous driving clips. It is notable for producing causal labels for autonomous vehicle data.
Show-Harness VLM Agent
Show-Harness is a harness for evaluating vision-language model agents on robot control tasks, with models and a dataset. It is notable as a fresh evaluation framework from Show Lab with high momentum.
VLAct VLA Pre-training
VLAct is the official code for representation-centric continued pre-training of vision-language-action models, with a continued-pretraining backbone and downstream checkpoints. It is notable for focusing on representation quality rather than data scaling for robot learning.
POLI Point Cloud
POLI is a self-supervised network that predicts a Gaussian ellipsoid per LiDAR point to capture local geometry on a statistical manifold. It is notable as RSS 2026 research code for improving 3D perception without labels or architecture changes.
OpenWAM World Model
OpenWAM is an open research stack for modular World-Action Model development, including infrastructure for composing model, representation, training, and evaluation choices, controlled studies, and a pretrained model. Notable for systematic world-action pretraining for generalist robot policies.
verl-vla Post-training Framework
verl-vla is a unified cloud-edge post-training framework for vision-language-action policies, built on verl, connecting distributed resources for human-in-the-loop data collection, fine-tuning, evaluation, and reinforcement learning. It is notable for providing a continuous post-training path across these stages.
WNM-3D Navigation Model
WNM-3D is the official implementation of a world navigation model with 3D scene conditioning for closed-loop vision-and-language navigation. It is notable as a 3D-conditioned navigation model for VLN tasks.