Pollen Robotics is taking an unusual approach with its newest machine: instead of asking buyers to trust a highlight reel, the Bordeaux-based team is shipping the entire training pipeline alongside the hardware. Microduck, a 25 cm bipedal robot now available for pre-order at $399, arrives with seven neural-policy behaviors already baked in — walking, sitting, kicking, roller-skating, and critically, getting back up after a fall. Every one of those behaviors was trained in simulation and exported directly to the robot.
The training stack is genuinely open. Pollen has published microduck_rl on GitHub, built on MuJoCo Warp with PPO, along with all the reward functions, domain-randomization settings, and the sim-to-real recipe. A usable gait trains in roughly one to two hours on a CUDA GPU at 4096 parallel environments. No local GPU? Appending —hf-jobs runs the same workload on Hugging Face Jobs. That’s a meaningful contrast to most sim-to-real research, where the actuator model and randomization choices stay locked in a lab notebook. Here they’re public.
The hardware is surprisingly complete for the price. Fifteen motors drive the legs, neck, head, and an articulated beak that can pick objects off the floor. Compute comes from a Rockchip RK3566 with 1 GB of RAM and 32 GB of storage. The sensor stack includes a front camera, two IMUs, an 8×8 time-of-flight LiDAR, microphones, a speaker, two NFC antennas, and Wi-Fi/Bluetooth. A removable NP-F550 battery delivers about an hour of runtime.
The detail that separates this from a toy is the actuator model. Pollen didn’t use an idealized PD controller for the Dynamixel XL330 servos. Instead, each joint runs the BAM M6 model — voltage control law, back-EMF, and Coulomb, Stribeck, and load-dependent friction. Randomization covers battery voltage, voltage sag, command delay, and ±1° of gear backlash. That’s the kind of engineering that determines whether a policy trained in simulation actually survives contact with the real world, and it’s rare to see it documented this thoroughly. The 61-dimensional observation space is shared across all policies, which lets the robot hot-swap between walking, recovery, and tricks mid-run. Reachy Mini, Pollen’s previous robot, sold over 10,000 units as a desktop manipulator. Microduck inverts that premise entirely: it’s built to fall and get back up.
