microduck

Product Introduction

1. Definition: The Pollen Robotics Microduck is a 25 cm tall, open-source bipedal robot designed specifically for accessible reinforcement learning (RL) and robotics research. It falls into the technical category of a low-cost, desktop-scale humanoid robot platform.
2. Core Value Proposition: It exists to solve the high cost and complexity barrier to real-world robotics and machine learning experimentation. Its core proposition is enabling simulation-to-reality (sim2real) transfer, allowing users to train AI behaviors in a physics simulator and deploy them directly to the affordable, physical hardware.

Main Features

1. Open-Source Software Stack: The entire software ecosystem, including the robot SDK, MuJoCo-based physics simulation environment, and the full reinforcement learning training stack, is publicly available on GitHub under a permissive Apache 2.0 license. This allows for complete transparency, customization, and community-driven development.
2. Pre-Trained Reinforcement Learning Policies: The robot ships with seven pre-trained locomotion and manipulation behaviors (e.g., walk, sit/stand, kick, get back up). These are not hard-coded scripts but are neural network policies that can be inspected, modified, and retrained by the user, providing a practical starting point for RL education.
3. Integrated Sim2Real Pipeline: The product is built around a validated workflow. Users train policies in the provided high-fidelity MuJoCo simulation (the robot's "digital twin") and can deploy them to the physical Microduck with minimal adjustment. The robot features a camera, LiDAR, and dual IMUs to facilitate this transfer and provide rich sensor data.
4. Modular Hardware & Accessories: Beyond the core robot, battery, and controller, Pollen offers optional packs. The Dev Pack includes spare motors and cables for durability and hacking, while the Accessory Pack (laser pointer, rollers, ball, NFC tags) extends the robot's interactive and training capabilities for advanced use cases.

Problems Solved

1. Pain Point: The extreme expense and technical overhead of deploying reinforcement learning research on physical robotic hardware. Traditional platforms cost thousands of dollars and require significant engineering effort for sim2real transfer.
2. Target Audience: Students and educators in robotics and AI; hobbyists and makers interested in hands-on machine learning; researchers and developers prototyping and benchmarking RL algorithms who need an affordable, reproducible physical testbed.
3. Use Cases: University and workshop courses on reinforcement learning; personal projects to learn robotics AI; prototyping and validating new locomotion policies for bipedal robots; creating interactive demos and artistic performances with a trainable robot.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros: Unprecedented price point ($399) for a capable, sensor-rich bipedal robot. Genuine end-to-end open-source model (hardware designs are not fully open, but software is). Functional out-of-the-box with entertaining pre-trained behaviors lowers the initial learning curve. Strong focus on a working sim2real pipeline, which is a major research hurdle.
   • Cons: As a small, lightweight robot (800g), its dynamics and payload capacity are limited compared to larger humanoids. The 25 cm size restricts the environments and tasks it can perform. Being a pre-order product (as of 2026), real-world durability and long-term community support are yet to be fully proven.

2. Key Alternatives & Differentiation:

   • Unitree Go1/A1: These are larger, more dynamic quadruped robots often used in research. They are significantly more expensive ($8k-$10k+) and are typically used for advanced mobility research rather than as an accessible educational tool for sim2real.
   • Robotis Darwin-Mini or OpenManipulator: These are popular in academia but often require extensive assembly and programming. They lack the integrated, pre-configured RL/sim2real stack that Microduck provides as its core offering.
   • DIY Biped Kits (e.g., based on Raspberry Pi & servos): While potentially cheaper, these require immense electrical, mechanical, and software integration work. The Microduck differentiates itself as a polished, fully integrated system where the core value is the ready-to-use training and deployment software workflow.

Frequently Asked Questions (FAQ)

1. What can you actually do with the Microduck robot? You can deploy and test pre-trained reinforcement learning policies for bipedal locomotion (walking, kicking, getting up), retrain these policies using the provided simulation, and develop entirely new behaviors like object interaction or skating using its accessories. It serves as a hands-on platform for learning and experimenting with machine learning in robotics.

2. Is the Microduck robot truly open source? The full software stack—including the SDK, simulation environment, and reinforcement learning training code—is open-source on GitHub under the Apache 2.0 license. The mechanical hardware designs and schematics are not fully open-source but are developed by Pollen Robotics.

3. How does the Microduck's sim2real transfer work? Policies are trained in a MuJoCo physics simulation that digitally twins the Microduck's dynamics. The robot's onboard computer runs these trained neural network policies at 50 Hz. The use of accurate simulation, domain randomization during training, and the robot's real-world sensors (IMU, camera) help bridge the "reality gap" for successful deployment.

4. What is included in the $399 Microduck pre-order? The base $399 package includes the complete 25 cm tall Microduck robot, a rechargeable battery, a USB-C cable for data/power, and a game controller for manual operation. Additional charger, development, and accessory packs are available for separate purchase.

5. Do I need a powerful computer to train policies for the Microduck? Training complex reinforcement learning policies can be computationally intensive. You can train on your own machine if it has a capable GPU or leverage cloud-based solutions like Hugging Face Jobs, which are integrated into the suggested workflow and supported via the Dev Pack credit.