A $399 bipedal duck learns its movements in simulation

Microduck is a 10-inch open robot whose movements can be retrained in simulation before being deployed to its onboard processor.

Standing 10 inches tall, weighing 1.8 pounds, and walking like a duck, Pollen Robotics’ latest machine uses a playful design to make reinforcement learning more tangible. Microduck arrives able to walk, stand up after falling, sit down, pick up small objects with its beak, and move on rollers. Its defining feature, however, is not performing those tricks but allowing users to modify them and create new ones.

Preorders open on August 27, 2026, at $399, excluding tax and shipping. Pollen Robotics plans to begin deliveries before Christmas 2026 but has not yet released a detailed schedule by country.

The box contains the robot, one battery, a USB-C cable, and a game controller. Microduck can therefore be used without starting a training run or writing code. The controller drives the robot, positions its head, lowers its body, and triggers its different movements.

Seven trained behaviors are announced for launch. The robot can walk at a requested speed, switch between sitting and standing, kick a small ball, lower its beak to the floor to retrieve an object, perform a forward roll, recover after a fall, and move when rollers are attached beneath its feet.

These actions are not part of a single fixed program. Each relies on a control policy that can be replaced with a new version. In this context, a policy is a small neural network that receives information about the robot’s state and determines which movements to send to its motors.

The onboard system runs on a Rockchip RK3566, a processor commonly found in small computers and connected devices. It runs ONNX-format policies locally without consulting a remote model for every movement.

The control loop operates at 50 Hz. Fifty times per second, the robot collects information including joint positions and body orientation, passes those values to the active policy, and applies the permitted commands to its motors.

Microduck has 15 servomotors connected to the same bus. Fourteen joints are directly represented in the training environment: five in each leg and four across the neck and head. The beak completes the assembly and can be controlled independently of the locomotion policy.

This structure allows the robot to move its head without interrupting its gait and to switch behaviors without resetting the entire control system. One policy can manage walking, another standing recovery, and a third a one-time movement such as a forward roll.

The policies share a common observation structure containing 61 values. Forty-eight describe the robot’s internal state. The remaining values represent movement instructions and requested poses for the head and body. This compatibility allows the onboard system to switch from one policy to another while the robot is operating.

The robot also includes a camera, motion sensors, and a depth-measurement system in its head. The documentation describes an 8-by-8 ToF grid capable of estimating the distance of objects in front of the beak. Pollen presents this equipment as a miniature LiDAR system.

These sensors do not automatically turn Microduck into an autonomous agent capable of understanding a room or selecting tasks on its own. The included movements are primarily triggered through the controller or a software command. The camera and depth sensor instead provide a foundation on which developers can build their own perception experiments.

Some demonstrations remain deliberately simple. The ball-kicking policy, for example, does not detect the ball’s position. It performs the expected movement when the user places the ball correctly and triggers the action. The videos therefore demonstrate mechanical capability and body control rather than a general understanding of the environment.

The central idea is to train movements inside a simulated copy of the robot. Pollen’s published environment uses MuJoCo Warp and PPO, a reinforcement-learning method in which the system tries numerous actions, receives rewards when it moves closer to its objective, and gradually adjusts its behavior.

Thousands of virtual Microducks can learn to walk in parallel. The command provided in the repository launches 4,096 environments simultaneously. Pollen estimates that one or two hours on a CUDA-compatible graphics card can produce an initial usable gait, without promising that it will immediately match the stability of the included policies.

Users without suitable hardware can send the training job to Hugging Face Jobs. This moves the computation to the cloud but may create charges depending on the machine and duration selected. The robot itself does not perform the heavy training workload: it runs the finished policy once that policy has been exported.

This distinction limits the immediacy of the phrase “teach it new tricks.” Showing the robot a movement or giving it a natural-language instruction is not enough. Creating a behavior requires defining the task, observations, objectives, rewards, and training conditions.

The simulator must also represent the real mechanism closely enough. A gait that remains stable in a virtual environment may fail on the physical robot because of its weight, friction, gear play, command delays, or declining battery voltage.

To reduce this gap, the environment varies physical properties during training. Voltage, power sag under load, delays, friction, and joint backlash can change between simulations. The policy therefore learns to work across multiple configurations instead of depending on perfectly consistent conditions.

Pollen also uses a detailed representation of the Dynamixel XL330 servomotors that accounts for voltage, back electromotive force, and several types of friction. At this scale, motor fidelity matters: small differences can be enough to destabilize a 1.8-pound biped.

This method improves the transition from simulation to physical hardware without guaranteeing it. A custom policy can produce unstable movement, make the robot fall, or place excessive strain on a joint. Progressive testing and physical