{"id":133403,"date":"2026-09-15T14:04:31","date_gmt":"2026-09-15T14:04:31","guid":{"rendered":"https:\/\/www.seeedstudio.com\/blog\/?p=133403"},"modified":"2026-09-15T14:04:35","modified_gmt":"2026-09-15T14:04:35","slug":"inspired-by-microduck-explore-open-source-embodied-ai-robots-and-physical-ai","status":"publish","type":"post","link":"https:\/\/www.seeedstudio.com\/blog\/2026\/09\/15\/inspired-by-microduck-explore-open-source-embodied-ai-robots-and-physical-ai\/","title":{"rendered":"Inspired by Microduck? Explore Open-Source Embodied AI Robots and Physical AI"},"content":{"rendered":"\n

MicroDuck has recently attracted a lot of attention from the robotics community. Its small form factor makes it approachable, but the technology behind it points to a much bigger shift: robots are increasingly moving from systems that simply execute programmed motions toward systems that can learn from data, perceive their surroundings, and adapt their behavior<\/strong>.<\/p>\n\n\n\n

\"MicroDuck<\/figure>\n\n\n\n

This shift brings together technologies that developers may have encountered separately: Robot Learning, Reinforcement Learning, Simulation, Computer Vision, LeRobot, Sim2Real, Edge AI, and Physical AI<\/strong>.<\/p>\n\n\n\n

The difficult part for someone just getting started is understanding how these pieces actually fit together.<\/p>\n\n\n\n

What does Jetson do in a robot-learning workflow? What does a robot-learning dataset contain? How does teleoperation become training data? Why train a robot in simulation if the final goal is to control a physical robot? And after a model has been trained, how does it actually produce movement?<\/p>\n\n\n\n

MicroDuck provides a good starting point because its reinforcement-learning workflow connects many of these pieces into one practical system.<\/p>\n\n\n\n

But you don’t have to wait for a MicroDuck to start experimenting.<\/p>\n\n\n\n

With platforms such as SO-ARM101, reBot Arm B601-RS, and LeKiwi<\/strong>, you can explore different parts of the same Physical AI stack today.<\/p>\n\n\n\n

This article follows that journey from teaching a robot a simple behavior to building systems that can perceive, learn, manipulate objects, move through an environment, and respond to natural-language commands.<\/p>\n\n\n\n

\n\n\n\n

Table of Contents<\/h2>\n\n\n\n
    \n
  1. MicroDuck: Where Robot Learning Starts<\/a><\/li>\n\n\n\n
  2. Your MicroDuck Hasn’t Arrived Yet? Start with a Robotic Arm<\/a><\/li>\n\n\n\n
  3. From Robot Learning to Real-World Manipulation: Meet reBot Arm<\/a><\/li>\n\n\n\n
  4. Why Simulation Matters: From MuJoCo and Isaac to Sim2Real<\/a><\/li>\n\n\n\n
  5. From a Robotics Tutorial to a Real Workflow: Seeed Workshop<\/a><\/li>\n\n\n\n
  6. Once You Understand One Robot, the Ecosystem Opens Up<\/a><\/li>\n\n\n\n
  7. From MicroDuck to Physical AI<\/a><\/li>\n\n\n\n
  8. Where Should You Start?<\/a><\/li>\n\n\n\n
  9. Explore the Seeed Studio Robotics Ecosystem<\/a><\/li>\n<\/ol>\n\n\n\n
    \n\n\n\n

    1. MicroDuck: Where Robot Learning Starts<\/h1>\n\n\n\n

    A traditional robotics program usually tells a robot exactly what to do. If you want a robot to walk, for example, you can manually define the position, timing, and trajectory of its joints.<\/p>\n\n\n\n

    Reinforcement learning takes a different approach. Instead of specifying every movement, you define a task and a way to evaluate the robot’s behavior. The robot then performs the task repeatedly, receives feedback, and gradually improves a policy that determines which actions it should take.<\/p>\n\n\n\n

    \"MicroDuck<\/figure>\n\n\n\n

    The MicroDuck workflow from Seeed Studio is a practical example of this approach.<\/p>\n\n\n\n

    What does Jetson actually do?<\/h3>\n\n\n\n

    The first piece is the NVIDIA Jetson Orin NX 16GB<\/strong>.<\/p>\n\n\n\n

    Jetson is not the algorithm that teaches MicroDuck to walk. It is the edge AI computing platform that provides the GPU resources needed to run the simulation, training, and later inference.<\/p>\n\n\n\n

    The workflow combines Jetson Orin NX, JetPack, CUDA, MuJoCo, PPO, and ONNX<\/strong>, but each technology has a different job.<\/p>\n\n\n\n

    Jetson<\/strong> provides the computing hardware.<\/p>\n\n\n\n

    CUDA<\/strong> provides GPU acceleration for workloads that benefit from parallel computation.<\/p>\n\n\n\n

    MuJoCo<\/strong> provides the physics simulation in which the robot can practice.<\/p>\n\n\n\n

    PPO (Proximal Policy Optimization)<\/strong> updates the robot’s policy during reinforcement learning.<\/p>\n\n\n\n

    ONNX<\/strong> provides a portable representation of the trained policy for inference and deployment.<\/p>\n\n\n\n

    Together, they form a development pipeline rather than simply a collection of AI keywords.<\/p>\n\n\n\n

    Where does MicroDuck practice?<\/h3>\n\n\n\n

    This is where MuJoCo<\/strong> becomes important.<\/p>\n\n\n\n

    Instead of placing a physical MicroDuck on the floor and asking it to learn by repeatedly falling, the workflow creates a simulated version of the robot.<\/p>\n\n\n\n

    The simulator handles the physics of the virtual environment, including gravity, contact, movement, and interactions between the robot and its surroundings.<\/p>\n\n\n\n

    The robot can then try a behavior, fail, reset, and try again.<\/p>\n\n\n\n

    For example, when training a walking policy, the simulated MicroDuck can repeatedly attempt to move forward while the system evaluates whether it maintains balance and follows the desired movement.<\/p>\n\n\n\n

    Because the environment is simulated, thousands of trials can be performed without physically moving or potentially damaging the robot.<\/p>\n\n\n\n

    How does PPO teach the robot?<\/h3>\n\n\n\n

    The training algorithm used in the workflow is PPO, or Proximal Policy Optimization<\/strong>.<\/p>\n\n\n\n

    You do not need to start with the mathematics behind PPO to understand what is happening.<\/p>\n\n\n\n

    At a practical level, the learning loop looks like this:<\/p>\n\n\n\n

    Take an action \u2192 observe the result \u2192 calculate a reward \u2192 update the policy \u2192 try again.<\/strong><\/p>\n\n\n\n

    If the robot performs behavior that matches the objective, the reward encourages the policy in that direction. If the robot falls or performs poorly, the reward provides a different signal.<\/p>\n\n\n\n

    After many iterations, the policy gradually becomes better at selecting actions that produce the desired behavior.<\/p>\n\n\n\n

    This is why the policy<\/strong> is the central output of reinforcement learning. It represents the behavior the robot has learned from its interaction with the simulated environment.<\/p>\n\n\n\n

    From training checkpoint to robot inference<\/h3>\n\n\n\n

    There is an important difference between training<\/strong> and deployment<\/strong>.<\/p>\n\n\n\n

    During training, the system produces checkpoints containing the information needed to continue or evaluate the training process. After training, the policy can be exported to ONNX<\/strong> for inference.<\/p>\n\n\n\n

    The workflow therefore looks roughly like:<\/p>\n\n\n\n

    Jetson \u2192 MuJoCo simulation \u2192 PPO training \u2192 trained policy \u2192 ONNX export \u2192 robot inference<\/strong><\/p>\n\n\n\n

    Once exported, the policy can be used to run behaviors on the robot, including examples such as walking, standing, sit\/stand transitions, ground picking, forward rolling, and ball kicking.<\/p>\n\n\n\n

    This is what makes the workflow interesting from a developer perspective. You are not simply downloading a robot demo. You can follow the path from building the training environment to creating a policy and eventually deploying that policy to hardware<\/strong>.<\/p>\n\n\n\n

    Explore the MicroDuck Reinforcement Learning on Jetson tutorial \u2192<\/strong><\/a><\/p>\n\n\n\n

    \"MicroDuck<\/figure>\n\n\n\n

    What if you want to teach MicroDuck something new?<\/h3>\n\n\n\n

    The official workflow also provides a path beyond predefined behaviors.<\/p>\n\n\n\n

    A new behavior can be divided into motion phases, with reward functions defining what the training process should consider successful.<\/p>\n\n\n\n

    For example, imagine that you want MicroDuck to perform a bowing motion. You could define the different phases of the movement, determine what posture should receive a positive reward, and specify when the behavior should be considered complete.<\/p>\n\n\n\n

    The workflow then becomes:<\/p>\n\n\n\n

    Define the task \u2192 configure the environment \u2192 train in simulation \u2192 visualize and evaluate the checkpoint \u2192 export the policy \u2192 deploy it \u2192 refine the behavior<\/strong><\/p>\n\n\n\n

    This is a useful way to think about reinforcement learning in robotics. The goal is not simply to understand an algorithm such as PPO. The goal is to build a repeatable process for turning a desired physical behavior into a trainable and deployable policy<\/strong>.<\/p>\n\n\n\n

    2. Your MicroDuck Hasn’t Arrived Yet? Start with a Robotic Arm<\/h1>\n\n\n\n

    If you are waiting for your MicroDuck, a robotic arm can be one of the easiest ways to start understanding Robot Learning.<\/p>\n\n\n\n

    A quadruped robot makes the learning problem exciting, but it also introduces balance, locomotion, and many degrees of freedom at the same time. A robotic arm gives you a more controlled environment in which the relationship between human demonstration, robot movement, data, and learning<\/strong> is easier to observe.<\/p>\n\n\n\n

    This is where SO-ARM101<\/strong> becomes a useful starting point.<\/p>\n\n\n\n

    \"SO-ARM101<\/figure>\n\n\n\n

    SO-ARM101 is an open-source robotic arm designed to work with the Hugging Face LeRobot<\/strong> ecosystem. It provides an accessible platform for experimenting with teleoperation, data collection, imitation learning, and simulation.<\/p>\n\n\n\n

    Seeed also provides resources connecting SO-ARM101 with LeRobot, Isaac Lab, and NVIDIA Isaac GR00T<\/strong>.<\/p>\n\n\n\n

    Explore SO-ARM101 + LeRobot + Isaac Lab + GR00T \u2192<\/strong><\/a><\/p>\n\n\n\n

    Start by making the robot follow you<\/h3>\n\n\n\n

    The first step does not need to involve an AI model.<\/p>\n\n\n\n

    After assembling and calibrating the system, you can use a Leader-Follower<\/strong> configuration for teleoperation.<\/p>\n\n\n\n

    You move the Leader Arm, and the Follower Arm reproduces the movement.<\/p>\n\n\n\n

    This simple interaction introduces an important concept in modern Robot Learning:<\/p>\n\n\n\n

    \n

    A human demonstration can become training data.<\/strong><\/p>\n<\/blockquote>\n\n\n\n

    Imagine that your task is to pick up a block and place it at another location.<\/p>\n\n\n\n

    Instead of writing a complete sequence of joint commands by hand, you can perform the task yourself while the system records what happened.<\/p>\n\n\n\n

    Repeat the task several times, and you now have multiple demonstrations of the same behavior.<\/p>\n\n\n\n

    \"SO-ARM101<\/figure>\n\n\n\n

    What is actually inside a robot-learning dataset?<\/h3>\n\n\n\n

    A robot-learning dataset is more than a video of someone controlling a robot.<\/p>\n\n\n\n

    Depending on the setup, an episode can contain synchronized information such as:<\/p>\n\n\n\n