Microduck

Product Introduction

1. Definition: The Microduck is a 25-centimeter tall, open-source bipedal robot designed for accessible sim-to-real reinforcement learning (RL). It is a compact humanoid platform equipped with 15 motors, a camera, LiDAR, and a functional grasping beak.
2. Core Value Proposition: It exists to democratize advanced robotics and AI research by providing an affordable, fully open-source hardware and software platform. Its primary value is enabling developers, researchers, and hobbyists to train robot behaviors in a MuJoCo physics simulation and deploy them directly to the physical robot, bridging the notoriously difficult sim-to-real gap.

Main Features

1. Open-Source Sim-to-Real Stack: The entire software stack, including the SDK, simulation environment, and reinforcement learning training pipelines, is open-source under the permissive Apache 2.0 license. How it works: Users train AI policies (behaviors) in a high-fidelity MuJoCo simulation on their local machine or via Hugging Face Jobs. The trained neural network policy is then flashed to the robot's onboard computer, where it runs at 50 Hz to control the 15 motors in real-time based on sensor input.
2. Pre-Trained & Retrainable Behaviors: The robot ships with seven pre-trained locomotion and manipulation policies, including a velocity-tracking walk, sit/stand, kick, grab, roller skating, and a get-up routine. Each behavior is a published, retrainable policy. Users can use these as starting points to fine-tune or completely retrain new skills using standard RL frameworks like Stable-Baselines3, leveraging the provided simulation.
3. Integrated Sensor Suite for Perception: The Microduck is equipped with a camera for visual input, a LiDAR sensor for distance measurement, and two Inertial Measurement Units (IMUs) for proprioception. This multi-modal sensing allows trained policies to react to the environment, enabling tasks like object interaction, navigation, and maintaining balance based on real-world feedback.

Problems Solved

1. Pain Point: The high cost and complexity of entry into physical reinforcement learning robotics. Traditional research platforms cost thousands of dollars and often have closed firmware, creating a significant barrier for students, indie researchers, and developers.
2. Target Audience: AI and machine learning students, robotics researchers focusing on sim-to-real transfer, hobbyist developers interested in embodied AI, and educators seeking a practical platform for teaching reinforcement learning and robotics.
3. Use Cases: Academic research in legged locomotion and manipulation; prototyping and testing RL algorithms with real-world physics; educational workshops on robotics and AI; developing and sharing novel robot behaviors (like dancing or playing soccer) within an open-source community.

Unique Advantages

1. Differentiation: Compared to other small robots (e.g., toy robots) it offers a serious, research-grade open-source stack. Compared to larger, more expensive research robots, it provides a focused, affordable, and accessible platform specifically for sim-to-real RL experimentation.
2. Key Innovation: Its end-to-end open-source approach and the careful co-design of its simulation model and hardware. The "simulated twin" is not an approximation but a direct digital counterpart, meticulously calibrated to ensure policies transfer reliably to the real 25cm robot with minimal real-world fine-tuning, solving a core challenge in robotics.

Frequently Asked Questions (FAQ)

1. What can you do with the Microduck robot? You can deploy its seven pre-trained behaviors out of the box, such as walking, kicking, and picking up objects. Crucially, you can retrain these or create entirely new behaviors using reinforcement learning in simulation and deploy them to the physical robot, enabling custom research and projects.
2. Is Microduck truly open source? Yes. Both its hardware designs and its complete software stack—including low-level firmware, control SDK, and the MuJoCo simulation environment—are open-source under the Apache 2.0 license, allowing for full customization, modification, and redistribution.
3. How does Microduck's sim-to-real training work? Behaviors are trained as neural network policies inside a precise MuJoCo physics simulation of the robot. Once trained, the same policy file is uploaded to the robot's onboard computer. The robot then runs this policy in real-time, using data from its camera, LiDAR, and IMUs to make control decisions for its 15 motors.
4. What is included in the $399 Microduck pre-order? The base package includes the 25cm robot, a battery, a USB-C cable for charging and data transfer, and a game controller for manual control. Additional packs for charging, development (spare parts), and accessories (like a laser pointer and rollers) are sold separately.
5. When will the Microduck robot ship? Pre-orders placed are scheduled to ship before Christmas 2026. The robot is being developed collaboratively by Pollen Robotics and Hugging Face.