Microduck Sim 2 Real

TL;DR

Pollen Robotics’s demo montage for Microduck, a ~800 g, ~25 cm tall open-source bipedal robot trained end-to-end in MuJoCo Warp via PPO on their Microduck RL — RL training environments for a ~800 g open-source bipedal robot stack. The video showcases the deployed real-robot behaviors — walking with velocity commands, fall recovery, standup from face-down / face-up / sitting poses, forward-roll (roulade), ball-kick, and roller-skating on passive wheels under the feet — all driven by ONNX-exported policies that share a 61-dim observation contract so the onboard runtime can hot-swap between them. No transcript is available for the video; the artifact of interest is the demonstration itself.

Key claims

The real robot executes at least six behavior classes from the underlying repo (walking, fall recovery, standup, sit/stand, ball-kick, forward-roll) plus a roller-skating variant on passive-wheel end-effectors [demo montage].
All demonstrated behaviors are ONNX policies produced by the sibling Microduck RL — RL training environments for a ~800 g open-source bipedal robot training stack — PPO on MuJoCo Warp with the BAM actuator model + domain randomization — and share the 61-dim observation contract that lets the runtime swap policies mid-execution [demo montage; see companion repo].

Method

The video is a real-robot demo montage without spoken narration or a transcript. The method being demonstrated is the mjlab + BAM-actuator + domain-randomization + ONNX-export pipeline documented in the Microduck RL — RL training environments for a ~800 g open-source bipedal robot README — see that paper page for the mechanical details.

Results

Qualitative demo footage only — no benchmark numbers are reported. The claim being validated visually is sim2real transfer works on the deployed robot for the trained task classes, not a quantitative headline.

Why it’s interesting

This is the “why should I care” companion piece to the Microduck RL — RL training environments for a ~800 g open-source bipedal robot release: the real-robot montage that shows the sim2real recipe actually closes on hardware. Interesting to a Luma reader because sub-1 kg bipedal sim2real is a regime where actuator fidelity dominates over kinematic accuracy — the BAM voltage-control-law actuator model + DR recipe visible in the demo is a much stronger prior than the ideal-PD model most humanoid stacks assume. Contrasts with the larger-scale humanoid open releases like AI Sapiens Open Source — ROBOTIS Docs Directory (ROBOTIS K1) that operate at a scale where kinematics dominates, and the demonstrated policy hot-swapping story (walk → recover → trick) is a concrete instance of the runtime pattern that Humanity's Last Machine: A Deep Dive on Humanoid Hardware identifies as an emerging humanoid-software-stack primitive.