Awesome Microduck

Microduck is a product of Pollen Robotics and Hugging Face. This is an independent, community-maintained list — not affiliated with, endorsed by, or sponsored by either company. Product names, logos and brands are the property of their respective owners, and every linked project belongs to its own authors under its own license.

Contents

• Ecosystem Status
• Official
• Documentation
• Simulation and Training
• Policies and Skills
• Datasets and Benchmarks
• Agent Tools and MCP
• Community Hubs and Registries
• Apps and Ports
• Hardware and Fabrication
• Articles and Coverage
• Videos
• Community
• Lineage

Ecosystem Status

The robot is pre-hardware: pre-orders are open, shipping is targeted for December 2026, and almost nobody outside Pollen has a physical unit yet. That shapes what is real today:

Simulation tools work now. The official browser sandbox, microduck_rl, and everything built on the published ONNX policies can be run and verified without a robot.

Hardware-facing tools are pre-validation. MCP servers, gateways and CLIs that target the real robotd API are written against the published design docs and mock transports; none listed here has been exercised on a shipped unit.

Policy distribution has landed. Pollen published the shipped policies to the Hub as microduck-policies on 31 August 2026, ten of them now, with velstand the default walk since 14 September. The policy channel (roadmap milestone M8) is on main: one .onnx per Hub repo beside a manifest.json stating kind, command encoding, observation and action sizes and robot compatibility, installed with robotctl policy list | load | add | reset | check | update | search, published from a checkpoint with uv run publish in microduck_rl, and browsable with an install button in the official Policy Playground. None of it has met shipped hardware yet.

Entries marked sim-only have not been validated on hardware.

Official

• pollen-robotics/microduck - The robot's software: robotd 50 Hz control loop, mediad camera/WebRTC, padd gamepad, updater, the official policy set and the robotctl policy channel that installs community policies from the Hub.
• pollen-robotics/microduck_rl - RL training environments on mjlab (MuJoCo Warp) with PPO, BAM actuator models, backlash simulation and domain randomization. Needs a CUDA GPU.
• pollen-robotics/microduck-gst-plugins - Prebuilt aarch64 GStreamer plugins (Rockchip MPP encoders, gst-plugins-rs WebRTC) used by the on-robot media daemon.
• microduck-policies - The ten official ONNX policies as a standalone Apache-2.0 repository on the Hugging Face Hub, described by one schema-2 manifest.json, with velstand the default walk since 14 September.
• Microduck Sandbox - Official in-browser simulator: MuJoCo compiled to WebAssembly plus onnxruntime-web running the real policies at 50 Hz, with gamepad support and the roller-skate variant.
• Microduck Console - Official hosted console for driving a duck that is not on your own network, over WebRTC and Hugging Face sign-in.
• Policy Playground - Official Space that reads every microduck- policy on the Hub the way the robot reads it and installs one onto your duck in one click, through the same calls robotctl policy add makes.
• Vision demo - Official Space that takes the duck's camera feed and runs the processing on Hugging Face hardware instead of on the robot.
• microduck-emotions - Official Apache-2.0 dataset of emotional body-language animations, published so the duck and Reachy Mini can act scenes together.
• microduck-duck-detector - Official one-class detector that finds other ducks in a duck's own camera, shipped as PyTorch, ONNX and INT8 RKNN for the robot's NPU and run on board by duck-detect. Scored at 0.80 mAP50, with every training run kept as a Git tag.
• pollen-robotics/duck_detector - The Apache-2.0 pipeline behind that model: capture, label, train and quantize, end to end.
• Arena reference policies - Pollen's baselines for the Microduck Arena challenges: a 2 m sprint, a 2 m roller sprint and a 1.5 m hill climb, each trained unchanged from its challenge recipe with the commit, seed and command needed to reproduce it. The challenge repository itself is not public yet.
• Product page - Specs, colorways and the launch story.
• Store - Pre-orders at $399.
• Press kit - Facts, full spec sheet, photos and downloads.
• Meet Microduck - Launch blog post from Pollen Robotics.

Documentation

All from the pollen-robotics/microduck repository.

• Docs index - Map of every doc in the repo.
• Cheat sheet - robotctl commands for day-to-day use of the robot.
• Architecture - How the daemons, control loop, policies and updater fit together.
• Design docs - robotd, updater, remote WebRTC, boot recovery, restart order and the WebRTC console.
• Policy channel design - How community policies reach robots from the Hub: slots, the origin rule, and the reasoning behind the manifest.
• Policy manifest - The schema-2 manifest.json contract: what each field means, and why only a constant-command episodic policy can become a one-shot skill.
• Roadmap - Milestones M1–M9, including the Hub model channel (M8) and autonomous brain (M9).
• duckctl - Controlling the robot from a laptop over Bluetooth.
• Pair a gamepad - Bonding a controller to the robot.
• Dev board setup - Setting up a development board and pushing branches.
• Contributing guide - Upstream contribution standards.

Simulation and Training

• microduck-rl-genesis - Genesis port of the walking task for AMD/ROCm GPUs, actuator model validated bit-exact against upstream. Sim-only.
• Isaac Lab Microduck port - Isaac Lab extension with Microduck assets, BAM/backlash actuators and RSL-RL tasks for walking, kicking and parkour. Sim-only.
• isaaclab-microduck (Newton) - Isaac Lab 3.0 (Newton MJWarp) port with locomotion, running, ball-kick and two-robot rally tasks, each A/B'd against the mjlab baseline. Sim-only.
• microduck-ros2-isaac - Tutorial for driving the public Microduck model from ROS 2 Jazzy and NVIDIA Isaac Sim, with joint control in RViz and the released walking policy in USD. Sim-only.
• microduck_description - ROS 2 description package: URDF and xacro generated from the public model, with 38 visual meshes and a robot_state_publisher launch file.
• microduck-ai-world - Embodied-AI playground pairing a robot runtime and RL policies with a detailed MuJoCo world and a vision-language brain that stays off the control loop. Sim-only.
• microduck-lab (DGX Spark) - Reproducible training and evaluation workspace for NVIDIA DGX Spark that pins the official runtime, simulator and microduck_rl as submodules.
• microduck-sim-playground - Lightweight educational workspace: bootstrap script, CPU MuJoCo viewer with keyboard poses, upstream pinned as submodules.
• microduck-rl-lab - Retrains five official skills and composes them into one automatic MuJoCo obstacle course. Sim-only.
• microduck-miniverse - Every published ONNX policy repackaged as a deterministic Miniverse simulation bundle.
• microduck-lab (Apple Silicon) - Trains policies on a Mac with no CUDA GPU, against the same MJCF model and 61/14 contract as upstream, with a live browser viewer. Sim-only.
• microduck-rl-on-thor - Getting the official training stack to run on aarch64 CUDA hardware (Thor, DGX Spark, Jetson), with every trap documented and verified on real machines. In English and Chinese.
• Microduck RL Ball Follow - MuJoCo Warp training repo adding a target-following task, a command-block contract layer and a drag-the-ball demo. Sim-only.
• Microduck School - Hosted Space where you set a lesson in plain English and watch the duck fail, retry and improve while the score climbs.
• MicroDuck Playground - Independent continuation of microduck_rl collecting reproducible experiments, policy demonstrations and printable hardware add-ons, rebased on a pinned upstream commit.
• Wicroduck - Attempt to put the whole loop behind a URL: MuJoCo compiled to WebAssembly steps the real MJCF in the browser with no Python and no backend. Simulation works today; in-browser training is the goal, not yet the state. Sim-only.
• isaaclab_microduck (PhysX) - Full port of the official training stack to IsaacLab and PhysX: 37 environments across walking, collision, roller, backlash and testbench models, BAM M6 actuator dynamics, PPO configs and ONNX export. The published checkpoint is an integration smoke test, not a converged gait. Sim-only.
• microduck-rl-torch - PyTorch-native rewrite of the training stack for machines without CUDA, with the same workflow carried up to an NVIDIA box or the cloud. Continuous integration and coverage on every commit. Sim-only.
• MicroDuck Unity Sim2Sim - Runs the official MJCF and ONNX policies inside Unity/Tuanjie with native MuJoCo kept as the physics authority and Barracuda doing inference, so the engine owns only scene, input and rendering. A Godot/Jolt counterpart now also runs the roller variant and a second robot in one window. Non-commercial use only, per the model license. Sim-only.
• MicroDuck Swan Lake - Trains a walking policy in the MotrixSim simulator, then blends it per joint with open-loop choreography (policy on the legs, choreography on the head and neck) for a two-minute ballet with a written root-cause log. In Chinese. Sim-only.
• duckbench - The physics bench under the golden vectors and the scored challenges: MuJoCo plus the shipped policies behind an HTTP service, a WebAssembly phone build, and the same bench exposed as MCP tools. Every published number names the plant it was measured in.
• mjlab-sycl - Companion package that gets mjlab's CUDA-only path training on an Intel integrated GPU or Arc card without editing your project: install it into the same environment and run one overlay command. Apache-2.0, with continuous integration. Sim-only.
• Microduck RL one-click toolkit - Two scripts that take a bare Linux box from no environment to a usable gait and an exported ONNX with its validation, timed at about four hours on an RTX 5060 Ti. In Chinese. Sim-only.
• microdux - JAX and MJX port of the training stack to MuJoCo Playground, with all fourteen official tasks registered for registry.load and the reward set checked term by term against upstream on a frozen transition. Sim-only.
• microduck-mjbatch - Trains on the CPU alone through mjbatch, thousands of MuJoCo instances stepped in a C++ thread pool: a walking policy solved in 2 hours 36 minutes on 8 vCPU with zero falls, and a ball-balance task reported as half-solved, with a median survival of 2.03 seconds and 78 percent of episodes still ending in a fall. Sim-only.
• RDK Robot Learning Platform - Vendor workbench from D-Robotics that takes a motion recorded in the browser simulator through training, evaluation and a pre-flight check toward its RDK-X5 board, all under one model contract. The local worker is a mock by default; the real training back end and board agent plug in through adapters. In Chinese.
• microduck_rl_unilab - Runs the upstream tasks on the UniLab distribution without depending on its source, pinned to upstream commit 29e887e with an alignment contract and an audit script that checks the port has not drifted from it. Apache-2.0, in Chinese. Sim-only.
• MotrixLab MicroDuck rollers - Passive-roller skating environment for MotrixLab built around alternating push-off cycles, heading control and recovery after contact, with a bundled checkpoint, a deterministic evaluation script and reward regression tests. In Chinese and English. Sim-only.
• microduck_genesis - Genesis World physics as a drop-in for the official stack: a TCP duck-body server the real daemons attach to with --sim, and an interactive viewer for rehearsing exported ONNX policies by keyboard. Apache-2.0. Sim-only.
• microduck_rl_tutorial - Beginner's Jupyter lab for omnidirectional velocity training on AMD ROCm, running from reinforcement-learning fundamentals through a smoke run, a train-from-scratch of about four minutes on an MI300X, an evaluation against a bundled reference checkpoint and keyboard teleoperation. MIT. Sim-only.
• microduck-locomotion-diagnosis - Reproduces the unanswered upstream issue #46, where the walking task converges to standing still, then finds three causes the reporter's seven experiments missed: the air-time reward pays out on whether a command exists rather than on whether the robot moves, lateral drift runs at twice the forward speed while both axes are weighted equally, and the headline metric mixes five terms. Gating the reward on achieved motion and weighting lateral error three times took error_vel_xy from 0.43 to 0.21. Also records what it took to run the CUDA-only stack on Windows. In Chinese. Sim-only.
• Microduck Carpet Lab - Configuration dashboard and training adapter for walking on compliant, variable-friction ground, laying out the observation and action contract, PPO settings and reward terms on a static site that can also recompute an imported evaluation report. States plainly that it ships no trained policy and that its contact model approximates carpet rather than simulating fibres. MIT. Sim-only.
• Microduck Academy (Hugging Face) - Type a sentence and an agent writes the reward and a judge, fine-tunes from a base checkpoint as short Hugging Face Jobs on your own account, has a vision model watch each take, then revises or ships a move repository with a daemon-ready policy.onnx and manifest. The harness, not the model, enforces a hard cap of 5 USD per session. Sim-only.
• mujoco-vecenv-cpp - Threaded C++17 environment whose claim is equality rather than speed: the same observations, reward and flags as its Python reference to the bit, tested on every commit, with NumPy's random generator reproduced exactly. It runs its parent project's PPO unchanged; training on it is the one result still open. MIT. Sim-only.
• Microduck RL WebUI - Web panel that launches training with editable PPO settings and plots reward and loss live, beside an inference pane streaming an off-screen MuJoCo render of a loaded ONNX policy under joystick, mouse or keyboard control. Apache-2.0, in Chinese. Sim-only.
• duckbatch - Batched, judged training on one 4 GB laptop GPU, where a pre-registered kill order prunes many small attempts and decision models are measured as judges rather than trusted as one. Its first result distils the default walker into a student 7.5 times smaller that tracks velocity within 5 percent of the teacher and gets up from a prone fall as often, but falls about three times as often under pushes. Apache-2.0. Sim-only.
• maxy_duck - The duck on a Windows PC: simulation running natively with the same BAM servo model and the official policies, training in WSL on the official stack, and a web monitor with live learning curves. Apache-2.0, in English and Chinese. Sim-only.
• microduck-lab (Colab) - Continuation of the Apple Silicon lab above that adds a Google Colab GPU training path for people without a local NVIDIA card, with a bilingual interface, crediting the original throughout. Apache-2.0, in Chinese and English. Sim-only.
• microduck-twin - Completes Pollen's own digital twin rather than replacing it: one command brings up the real daemons against the official simulated body, then adds what that setup leaves out, an audible voice, battery and thermals that follow what the robot is doing, the head microphone as a sensor, a console and the roller variant. No policy is retrained. Apache-2.0. Sim-only.
• microduck_rl_mac - The official training environments on Apple Silicon with the task managers, rewards, randomisation and BAM actuator model left intact, running physics on the CPU and PPO on the Apple GPU. Flat-ground walking is demonstrated on an M1 Max. Apache-2.0. Sim-only.

Policies and Skills

Community-trained policies and task definitions built on microduck_rl. This section also publishes a versioned machine-readable registry for MicroDuck Lab and DuckHub. Known unsafe, compromised, removed, or superseded entries are published separately in the machine-readable revocation registry. Consumers should resolve both files from the same verified Git commit and stop a revoked policy before execution. Submit a policy with the structured community form.

• microduck-backflip - Reproducible mjlab backflip task with an evaluation battery, experiment log and explicit safety gates. Sim-only.
• microduck-courier - Pick-carry-place task in an apartment scene with a trained policy, rollout clip and telemetry. Sim-only.
• Jump playground - Browser sandbox fork with a custom-trained vertical-jump policy; live demo. Sim-only.
• microduck-sidekick-dance - Drop-in mjlab task for a lateral dance step, with reward design notes. Task only, not yet trained.
• microduck-step-up-policy - Policy pair for crossing a 25 mm square-edged step using the head as a temporary brake, then recovering upright; weights on the Hub. Sim-only.
• microduck-max-height-jump - Deployable ONNX export and reproducibility evidence for a one-shot PPO jump, with the height claim stated as a training objective rather than a physical maximum. Sim-only.
• microduck-walking yaw ablation - Single-variable reward ablation that cuts yaw-rate variance against the released alpha_walking baseline, with the evaluation harness that makes the comparison reproducible. Sim-only.
• Microduck Circus - Three ducks learn to skip a shared long rope, two turning and one jumping, through an act-fail-practice-adapt agent loop with a holdout round. Sim-only.
• microduck-flamingo-cycle - One-legged flamingo pose policy on the Hugging Face Hub.
• microduck-polite-bow - Bow gesture policy on the Hugging Face Hub.
• microduck-moonwalk-backward - Backward moonwalk gait on the Hugging Face Hub.
• More policies on the Hub - The growing long tail of community-trained gaits and gestures, searchable on the Hugging Face Hub.
• MicroDuck skills for TorchRL - One recurrent task-conditioned policy covering seven skills (stand, four directions, turn each way) through a task embedding and a 128-unit GRU, trained on 12 million transitions across 16 native CPU environments, with all 56 evaluation episodes surviving the full horizon. Published by the TorchRL project, which also states its known limits: the head still drops about 63 degrees walking backward, and turns reach roughly two thirds of the commanded rate. Sim-only.
• microduck-basketball - Balances on a free-rolling size-7 basketball and follows velocity commands from proprioception alone, through a one-layer LSTM with no ball state in the actor inputs. Simulation and export checks pass; onboard timing and real-robot behavior are untested.
• microduck-swing - Starts motionless at the bottom of a two-cord swing and pumps itself with its head and legs to a 173 degree span, with 71 of 100 randomized seeds clearing every full-horizon validity gate and rollouts rejected for cord slack, lateral drift or attachment misalignment rather than judged on angle alone. Printable seat and straps included. Sim-only.
• microduck-stilts - Eight forward-walking policies for stilts from 10 cm to 2 metres, each height its own ONNX graph, continuation checkpoint, manifest and video behind a machine-readable index, with the printable stilt hardware alongside. Sim-only.
• microduck-running - Running policy robustified from a 1.687 m/s speed frontier with velocity pushes, centre-of-mass variation and initial tilt, holding 1.651 m/s nominal and 1.612 m/s under a combined backlash and disturbance battery. The author calls it a hardware candidate and says heading and lateral drift remain substantial. Sim-only.
• microduck-climb - A climbing policy and a get-up policy for a desk ladder, with full checkpoints, hashes and printable ladder parts. The switch between them reads simulator contact rather than an onboard detector, and the ladder parts are geometry-checked but not yet printed or load-tested. A chimney variant climbs a narrow corridor by bracing between two walls. Sim-only.
• Raised-platform backflip and long jump - Two jump policies from a raised platform, a backflip dropping 0.8 m onto a soft-contact mat and a long jump across a 30 cm gap down to a lower platform, each with a checkpoint, an ONNX export and a config naming the command slots. The backflip card warns that landings may break the robot and that a soft simulated mat says nothing about a rigid floor. Apache-2.0. Sim-only.
• microduck-happy-hop - Short two-foot hop that crouches, takes off, absorbs the landing and returns home, and one of the first community policies shown running on a physical MicroDuck, after Pollen restored the one-step action delay it depends on. The card states that delay alongside the rest of the contract.
• beak-throw - Winds up, throws a 24 mm ball from the beak and recovers to a stand as a one-shot skill, with weights on the Hub. It completed 50 of 50 randomized trials only with a mandatory runtime clamp and failed the stricter raw-output gate, and a safety note governs any first hardware run. Sim-only.
• microduck-rl running - Training repository with 39 registered tasks and a published straight-running policy holding 1.844 m/s over 10 seconds and 1.806 m/s over 30 under a heading-hold controller, stating plainly that the 2.0 m/s target was not met. Sim-only.
• rough-walk-e - Drop-in replacement for alpha_walking fine-tuned on a hostile-terrain ladder of tiles, small stairs, rubble and slopes, with roughly half the falls; the from-scratch rough-walk-g is sturdier on stairs and slopes at 17 percent more motor power. Stated limits include falling on downward steps of 2 cm or more. Sim-only.
• Collision Flamingo II - One-foot balance trained against a full exterior-shell collision model, shipped with the exact overlay needed to reproduce it. It works through the ONNX deployment rehearsal but not through the training repository's own play viewers, and the card reports that discrepancy as unresolved. Sim-only.
• MicroDuck Jump - Raises happy-hop's band to an in-place hop of 3 to 6 cm on stock servos, with the checkpoint chosen by in-band rate rather than by recency: over 128 episodes half land inside the band and none fall. Its own acceptance checklist records the failure as well, a 128 N landing peak against a 40 N target, with the fix identified but not yet validated. Sim-only.
• microduck-standup - Episodic get-up that runs for six seconds and returns itself to a standing pose, published against manifest schema 2 with its robotctl policy add line and its training branch and commit named. The card also admits the export came from a checkout with uncommitted changes. Sim-only.
• microduck-groove - A ten-second in-place dance of rhythmic knee bounce and a small head sway that returns to standing, published with its robotctl policy add line. In simulation it never falls, with 1.6 degrees of trunk tilt at most, and it bounces at 87 percent of the reference amplitude and about 126 BPM against a 100 BPM target; the planned head nod, tilt and hip sway were not learned. The training fork is unpublished, so its commit does not exist upstream. Sim-only.
• Microduck headstand - Six policies chained into an acrobatic routine, folding, kicking up with legs together, rolling back, entering a split headstand, switching the leading leg and splitting over to standing, with the official standing policy between moves. The full routine passed 87 of 96 attempts across three seeds and each policy 32 of 32, with raw reports, checkpoints and normalised ONNX exports on the Hub and the training tasks proposed upstream. Sim-only.
• Microduck Vision Soccer - Head-camera striker against a monocular goalkeeper, built so the same perception and state machine can run on the robot through mediad and robotd. Its frozen ground-truth navigation baseline scored 98 of 100 goals with no falls, reported separately from the camera-driven controller on purpose; full soccer has not been run on hardware. Apache-2.0. Sim-only.
• microduck-rl (CPU) - Balance-and-recover policy trained on a laptop with no GPU in plain CPU MuJoCo, evaluated on physics models the robot's own authors supplied rather than on its own randomisation. The first 25M-step training chunk beats a PD baseline, and the repository reports that its domain randomisation, as run, made standing worse. Its tests also catch a silent indexing bug that is wrong on 13 of 14 joints on the backlash model. MIT. Sim-only.
• MicroDuck Piano Learning - Two ducks walking across a passive floor keyboard, published as an honest snapshot: one accepted traversal in which 1 of 64 evaluation worlds passed every strict gate, pressing and releasing six keys without a fall. Its author is clear it is not yet a score-following performance. Apache-2.0. Sim-only.
• microduck-tricks - Three skills trained on Hugging Face Jobs and scored on held-out batteries against the shipped policies: a run into a forward roll that hands off from walking 200 times out of 200 where the official policy manages 86 percent, a cone slalom that clears 8 of 8 gates without a fall where the official policy clears 21 percent, and a one-foot skate that lifts a blade for 0.4 to 0.6 seconds of true single support and skates on, 24 takes without a fall. A sustained one-blade glide is reported as unreachable on this robot, with the measured ceiling. Sim-only.
• microduck-onefoot-skate - Interim checkpoint on a sustained left-foot glide, published as method and measurements rather than as a result: the declared 2.0 second gate is cleared by only 1 episode in 64 and no stage has passed formal acceptance, but 96.9 percent hold a credible one-foot glide past 1.0 second and the best recorded is 2.14 seconds. Carries the learnable-authority audit that found 35 percent of every episode had nothing to learn, two spawn-state defects traced to wheel velocities in the observation, all 15 training runs as curves, signed decision records, and an ONNX path documented well enough to drive the policies without installing MuJoCo. Sim-only.
• microduck-detector - YOLO11n detector that finds a Microduck in an image, 2.6M parameters, scored at 0.63 mAP50 on a held-out split of synthetic renders and real press photos; try it in what-the-microduck.
• Microduck RL 4096x6000 - Velocity-tracking reproduction pinned to an upstream commit: 4096 parallel environments, 6000 PPO iterations, with intermediate checkpoints, the ONNX export, training config, TensorBoard events and closed-loop replay video all kept, and a matching ball-kick run keeps the same record. Sim-only.

Datasets and Benchmarks

Fixtures and scored tasks for checking a runner or a policy against numbers rather than against a video.

• Policy golden vectors - Observation and action pairs recorded from the shipped policies, so an independent runner can be checked against the same numbers. A conformance fixture, not weights.
• Microduck Ball Challenge - Scored ball-chasing benchmark with the physics plant pinned by hash; a stairs challenge follows the same discipline. Sim-only.
• Trajectory dataset - Multi-modal state-action trajectories from the 14-DOF simulated robot, aimed at offline reinforcement learning and imitation. Sim-only.
• Microduck detection dataset - Labelled bounding boxes over synthetic renders, composites and real press photographs, the training and validation split behind the detector above.
• Microduck skill tree - Build log for a ladder of skills trained against a pinned upstream commit on one GPU, one folder per level whether it shipped or not, with the installable policies in their own repositories such as walk-recover. It reports its misses too, like a recovery policy that failed to get up from its back in 3 of 3 takes. MIT. Sim-only.
• Hermes Agent runs - One sentence typed to Hermes Agent and the duck carries it out, with every run logged: the sentence, the agent's full transcript, the simulator's ground truth, the robot daemon's own monitor and a replay. Only the body is simulated; robotd, its safety layer and its control loop are the real ones. Apache-2.0. Sim-only.
• Microduck community feedback - Human judgments collected from a planning studio: corrections to how a plain-language router labelled each instruction, and pairwise preferences between two trained walkers doing the same thing. CC0.

Agent Tools and MCP

Ways for LLM agents and scripts to drive a duck.

• joeynyc/microduck-mcp - Agent-agnostic MCP server modeled on the real-robot robotd architecture: mock, MuJoCo sim, Unix-socket and SSH transports behind one tool set, with a safety layer. Hardware transports are pre-validation.
• aj-dev-smith/microduck-mcp - MCP server driving the simulated duck (CPU MuJoCo with the official policies), with rendered camera frames as tool output and a live agent-experience debug page. Sim-only.
• quackd - Agent CLI that began as the brain daemon the duck was missing, named after its own services: register robots once, then state a goal in plain language and Claude, OpenAI, Gemini or Grok sequences each robot's existing skills, alone or as a flock. .duck task files, safety rules, MCP support, on PyPI. It now drives seven robots, all in simulators or mocks so far.
• meckie-duck-gateway - Holds the robot's WebRTC/JSON-RPC session and re-exposes it as a small local HTTP API, with a hardware-free protocol double for testing.
• OpenCastor Microduck integration - Third-party robot-agent framework that discovers Microducks, sends intents through robotd and composes routines.
• Strands Robots Microduck provider - Python/MuJoCo provider wrapping the shipped ONNX policies behind a common simulation and hardware interface.
• Microduck Lab (gr.Workflow) - Hosted Space that turns a plain-English routine into a sequence of the robot's skills through a language model, then plays it back.
• quacksat - Turns the duck into a roaming voice satellite, with interchangeable Home Assistant Wyoming, agent-bridge and direct backends selected from one config file.
• MicroDuck TinyVLA - Vision-language-action model driving a live MuJoCo duck from a head-camera frame, the 61-float state and a plain-English instruction, all on ONNX Runtime CPU. Sim-only.
• microduck-cli - Same verbs for an agent and for a human: five noun groups over robotd's own JSON-RPC socket, --json on every command, and an empty runtime dependency list. On PyPI; checked against the real daemon and the MuJoCo body.
• microduck-plugin - Agent Plugins package that gives Cursor, Copilot, Codex, VS Code or Kiro the train, export, publish-to-Hub and install loop as a single skill. MIT, with continuous integration.
• microduck_demo_part1 - Drives the official walking policy in MuJoCo over MQTT through a small adapter, with upstream pinned as submodules and a one-script bootstrap; motor control and inference are left untouched. Sim-only.
• microduck-remote-policy-server - Experimental server for running policy inference off the robot at 50 Hz: it checks each ONNX against the 61/14 contract and stays silent for a network it did not load, so robotd falls back to its local copy. Pairs with an experimental branch of a runtime fork.
• Harboria - Rust runtime for an assistant that holds long-term context and decides when to intervene, rather than answering one request at a time, with the duck as its first adapter. Apache-2.0 and early: the domain model and daemon are there, the robot side is not yet exercised on hardware.
• Jev drives a MicroDuck - Replaces the gamepad with a typed-judgment model that sees only the head camera rendered as characters and picks the next move about every 150 ms, leaving the 50 Hz locomotion policy untouched. Sim-only.
• Jevduck - Browser research simulator asking whether a model can choose useful next actions from measured robot state while a separate controller keeps execution bounded, and keeping an accepted instruction and the physical result it actually produced as separate events. MuJoCo WebAssembly with the official policies, a live build and a validation record. Sim-only.
• quack-nav - Navigation daemon beside the released robotd: time-of-flight mapping and relocalization, a cliff guard, a costmap planner, named places, an explorer that maps a house a charge at a time and go_to, as JSON-RPC tools, with its twin results published, misses included. Apache-2.0, in English and Italian. Sim-only.
• Laya Vision find-and-kick - A 201M vision-language model fine-tuned to drive the duck from its own head camera, choosing forward, left, right or kick about every 70 ms to find a ball and kick it in quackd's MuJoCo simulator, with a live demo that plays whole episodes. CC BY-NC-SA 4.0. Sim-only.

Community Hubs and Registries

• MicroduckHub - Community policy browser listing the shipped behaviors and Hub retrains, with one-click deploy planned for when upstream's model channel ships.
• uDuck Registry - Independent catalog of community policy descriptors with contract specs and verification status.

Apps and Ports

• DuckKit - The Microduck as a pure Swift package: runs the real ONNX policies with the real 61-float observation at 50 Hz, plus kinematics, protocol types and Linux tests.
• microduck-sim (iPhone) - The nine shipped policies plus MuJoCo running natively on iPhone in Swift/RealityKit, with an AR mode at true scale.
• Microduck WebXR - Physics-driven Microduck for Meta Quest and ordinary browsers, with controller walking, learned skills, room placement and an autonomous wander mode.
• Microduck Anatomy - Interactive holographic anatomy viewer with staged component focus and exploded assembly views.
• microduck-tracking - Multi-object tracking on top of roboflow/trackers that gives the duck a target lock, so it fetches one thrown ball past identical distractors.
• Microduck AR - WebXR adaptation of the official sandbox with AR placement and ground-pick interaction.
• specs-microduck - Hand-gesture teleoperation of the simulated duck from Snap Spectacles, with in-lens telemetry.
• MicroDuckModels - Browser simulator rebuilt on Three.js and React Three Fiber with MuJoCo WebAssembly physics and local ONNX inference, running all nine shipped policies. Readme in Chinese and English.
• microquack - Procedural droid-voice synthesis for the duck: a Rust core rendered live in the browser via WebAssembly, also on Hugging Face.
• RL Physics Overlay - Dependency-free telemetry overlay for the browser simulator showing joint forces, torques, contacts and learning signals without blocking the training loop. In Chinese and English.
• Microduck Studio - Local control room that puts robotd status, safe control and the MuJoCo body behind one browser page, deliberately duplicating none of the safety, inference or physics it fronts. Readme in English and Chinese.
• Kinematic viewer - Drag any joint through its real range and watch the chain follow, with axis, hard limits, trainable limits and home angle shown live. One index.html, no build step; also a hosted Space.
• 3D bipedal teleop - Browser digital twin driven by the shipped policies over ONNX Runtime, with omnidirectional teleoperation and reported velocity-tracking error. In English and Korean. Sim-only.
• esp-duck - All nine shipped policies compiled onto an ESP32-S3 and hot-swappable at runtime, holding 50 Hz with roughly two times margin on per-channel INT8 weights and FP32 activations, fed by the board's own IMU, with a quantization-accuracy gate that runs on the host, on the device and at every boot.
• spacemit-microduck - Brain swap to SpacemiT RISC-V: the official Rust runtime cross-compiled natively for K3 and K1 boards driving the servo bus directly, training left on the existing x86 and GPU chain. K3 measured; K1 still in progress. In Chinese.
• microduck-arena (3v3 football) - Three-a-side reinforcement-learning football in the browser on MuJoCo WebAssembly and onnxruntime-web, with a playable site. MIT, readme in English and Chinese. Sim-only.
• microduck-viewer - Runs a trained locomotion policy as plain JavaScript at 50 Hz with no inference runtime at all, against MuJoCo WebAssembly physics, as embeddable pages for an interactive control textbook. No server, no build step. Sim-only.
• Microduck Academy (classroom) - Staged reinforcement-learning classroom for software engineers new to the field: read the task, write Python, run real tests and watch the result, all on one page, ending in the official 3D simulator. Independent and explicit about it. In Chinese and English.
• microduck-docker - The simulator, MuJoCo physics and the pretrained policies in a single published container, so trying the robot is one docker run rather than an environment. Apache-2.0, built in CI and on Docker Hub.
• DuckFly - Wires a simulated 668-neuron fly circuit to the duck: map any of 19 circuit signals to 10 robot actions in a wizard, then watch the neural activity and the resulting movement side by side. Runs locally in a browser or as a Mac app, with a hosted version. Sim-only.
• FlyWire MicroDuck - Couples a fixed spiking model of the whole released fly connectome, 139,255 neurons, to a trained walking policy; only an 84k-parameter readout from descending neurons is learned. Reaches 95 of 100 held-out goals without a fall, against 0 of 100 with the stimulation, the readout features or their identities removed. Sim-only.
• Microduck Color Studio - Colors and materials per part across the 70 instances of the simulation assembly, then exports a color-preserving multi-plate 3MF and per-part STLs for printing. Apache-2.0 code with the model's own provenance kept in its notice, and a live version. In Chinese and English.
• Microduck Racer - Four roller-skating ducks race waypoints around a furnished dining room, a racing layer on top of the unchanged roller policy and observation contract. Sim-only.
• Microduck Playground Competition - Pick the winner of a two-lap race between three ducks, each running its own physics and policy in the browser, and win or lose fish on the result. Apache-2.0. Sim-only.
• Duck-Man - Maze tag for five ducks walking on the published gait, with a trained stand-up policy for tagged or fallen ducks and a strategy network evolved on a laptop CPU. Apache-2.0. Sim-only.
• Bohemian Rhapsody by the Microducks