Microduck: A Detailed Look at Open-Source Bipedal Robot Sim2Real in Practice

Microduck: An Underestimated Open-Source Milestone

While Chinese robotics companies release jaw-dropping videos of kung fu and dancing robots on a near-weekly basis, a small bipedal robot called Microduck might not seem particularly impressive at first glance. But according to a Pollen Robotics engineer sharing on Reddit, this robot's significance goes far beyond being a "toy" — it's a fully functional bipedal humanoid robot equipped with an open-source software stack and a sim2real (simulation to reality) pipeline that works surprisingly well.

Pollen Robotics is a French robotics company previously known for its open-source humanoid robot Reachy — an upper-body humanoid featuring biomimetic arms and an expressive head, primarily designed for human-robot interaction and service-oriented research. Microduck can be seen as Pollen Robotics' pivotal step toward full-body humanoid robots, particularly bipedal locomotion, continuing the company's longstanding commitment to open source and community collaboration.

The engineer, who has over a decade of experience in robotics competitions, put it plainly: "Getting a bipedal robot to stand, walk, and even rollerblade on its own is no small feat." It's precisely this frontline experience that gives him a clear-eyed appreciation of the project's value. In his view, enabling thousands of hobbyists to experiment with reinforcement learning on a bipedal robot is nothing short of "crazy" — if a platform like this had existed when he started out in 2010, it would have saved him years of trial and error.

What Is Sim2Real: The Core Challenge of Bridging Simulation and Reality

The Reality Gap

Sim2real (Simulation to Reality) is one of the most critical and thorny challenges in robot learning. Researchers typically train locomotion policies in physics simulation environments — because simulation can generate massive amounts of training data at low cost, high speed, and in parallel, making it especially well-suited for reinforcement learning (RL) algorithms that require extensive trial and error.

Reinforcement learning is a machine learning paradigm where an agent learns optimal strategies through repeated interaction with an environment. Unlike supervised learning, RL doesn't require pre-labeled "correct answers" — instead, it uses reward signals to guide the learning process. In robotics, RL is particularly well-suited for solving motion control problems: it's extremely difficult for human engineers to manually write control rules that allow a bipedal robot to walk stably across various terrains, but RL algorithms can automatically discover effective gait strategies through millions of virtual trials. In recent years, with advances in GPU computing power and parallel simulation technologies (such as NVIDIA Isaac Gym), RL applications in robotics have moved from the lab to real-world deployment. Teams at DeepMind, Tesla Optimus, Unitree, and others have all adopted RL-driven locomotion control approaches.

But here's the problem: there is a gap between the simulated world and the real world that's extremely difficult to bridge. Motor response delays, joint friction, sensor noise, floor surface differences… even slight deviations in these details can cause a policy trained to perfection in simulation to immediately "crash and burn" when deployed on real hardware. This is what's known as the "Reality Gap."

Mainstream approaches to addressing the Reality Gap include: Domain Randomization, which introduces large-scale random perturbations to physics parameters, visual appearance, and other variables during simulation training, forcing the policy to adapt to various uncertainties; System Identification, which involves experimentally measuring the precise physical parameters of real hardware and feeding them back into the simulation model; and combinations of both. Microduck's approach leans more toward the latter — reducing the gap between simulation and reality at the source through precise actuator modeling.

How Microduck Solves the Sim2Real Transfer Problem

The key reason Microduck's sim2real pipeline "works surprisingly well" lies in its high-quality actuator modeling. The engineer specifically credited the Rhoban team's contributions, along with their aptly named open-source repository — BAM (github.com/Rhoban/bam).

Rhoban is a robotics research team at the University of Bordeaux, France. They've won multiple RoboCup robot soccer championships and have accumulated deep expertise in locomotion control for small humanoid robots. BAM (a name that carries a satisfying "nailed it" kind of humor) is their open-source actuator precision modeling tool. It can perform systematic calibration and modeling of common servo motors — including torque-velocity curves, control delays, gear backlash, thermal characteristics, and other physical parameters that are usually overlooked but are critically important for sim2real transfer.

The precision of the actuator model directly determines whether motor behavior in simulation can faithfully reproduce real motor behavior. Traditional approaches often use idealized motor models that assume a linear relationship between torque and current and ignore nonlinear friction and backlash in gear transmissions. These "minor details" are precisely the culprits behind sim2real failures in tasks like bipedal walking that demand precise force control and rapid response. BAM builds a more realistic digital twin of the actuator through experimental calibration, ensuring that the servo torque, speed, and delay curves in simulation closely match reality. Only then can the trained walking and balance policies successfully transfer to the physical robot. In essence, the precise actuator model provided by BAM is the invisible cornerstone that makes this entire sim2real pipeline viable.

Open-Source Value: Making Bipedal Robot Reinforcement Learning Accessible

In the past, doing reinforcement learning on bipedal robots was an exclusive game for top-tier labs — you needed expensive hardware (a research-grade humanoid robot can easily cost hundreds of thousands of dollars), complex software stacks (spanning simulation engines, RL frameworks, communication middleware, real-time control systems, and more), and deep engineering expertise. Microduck open-sources and democratizes this entire capability.

For robotics hobbyists, students, and independent researchers, this means:

• No need to build simulation environments and hardware drivers from scratch;
• The ability to experiment with your own RL algorithms directly on a proven, working platform;
• Real hands-on experience with the complete loop of "train a policy → deploy to hardware → observe walking performance."

The practical value of this "complete loop" should not be underestimated. Many RL researchers spend their entire careers in pure simulation, never having the opportunity to deploy policies on real hardware, and consequently lack intuition about the many practical challenges of sim2real. Microduck fills exactly this gap — it makes the "last mile" from paper to physical robot no longer an insurmountable chasm.

This kind of low-barrier hands-on opportunity is invaluable for cultivating the next generation of robotics researchers. As the original author reflected, this is something he desperately wanted but couldn't have when he entered the field in 2010.

Future Directions: From Bipedal Walking to Autonomous Intelligent Behavior

The engineer candidly noted that Microduck's locomotion capabilities are "already quite good, but there's still room for improvement." The real next big challenge is moving toward autonomous behaviors.

He outlined several exciting directions for exploration:

• Natural Language Understanding: Enabling robots to understand human instructions. With the rapidly advancing capabilities of large language models (LLMs), researchers are exploring how to integrate models like GPT and LLaMA into robot systems, allowing them to translate natural language commands like "go to the kitchen and get me a glass of water" into specific action sequences. The core challenge here is aligning semantic understanding of language with the constraints of the physical world — the robot must not only "understand" the instruction, but also comprehend what executing that instruction means in the current physical environment.

• SLAM (Simultaneous Localization and Mapping): Enabling robots to perceive and build a map of their environment. The core problem SLAM solves is: how can a robot in a completely unknown environment simultaneously build a map while determining its own position within that map? This is a classic chicken-and-egg problem — accurate localization requires a map, but building a map requires accurate position information. Modern SLAM solutions typically fuse data from multiple sensors including LiDAR, cameras, and IMUs, using probabilistic filtering or graph optimization algorithms to solve both simultaneously. For bipedal robots, SLAM faces an additional challenge: the intense shaking during walking degrades sensor data quality, requiring more robust algorithm designs.

• Autonomous Navigation: Self-directed path planning and movement through space;

• Grasping: Physical manipulation-level interaction with the real world. Dexterous Manipulation is another major frontier in robotics, involving key technologies such as force control, tactile sensing, and fine-grained motion planning.

Stacking these capabilities on top of the existing bipedal locomotion foundation will transform Microduck from a "locomotion platform" into an "agent platform." This is precisely the central proposition of Embodied AI research today — truly combining the cognitive capabilities of large models with the physical execution capabilities of robots.

The core idea behind Embodied AI is that true intelligence cannot exist independently of a physical body and environmental interaction. Unlike large language models that operate purely in the digital world, embodied intelligence requires AI to perceive the three-dimensional physical world, understand the physical properties of objects, and execute actions through a body to complete tasks. This research paradigm is giving rise to a "brain + cerebellum" technical architecture: large language models or vision-language models serve as the "brain" handling high-level task planning and semantic understanding, while RL-trained locomotion policies serve as the "cerebellum" handling low-level motor control. The value of open-source platforms like Microduck is that they provide researchers with a physical vehicle for truly validating this "cognition-action" loop, rather than merely drawing up plans in simulation.

Conclusion: The Open-Source Robotics Community Needs More Microducks

Microduck may never go viral like those kung fu robots, but it represents a more humble and more enduring kind of value: making cutting-edge technology accessible to everyone willing to get their hands dirty.

A sim2real pipeline that actually works end-to-end, an open-source software stack, plus the carefully crafted actuator models from the Rhoban team — together, these form a reproducible, learnable, and extensible bipedal robot platform. For the open-source robotics community as a whole, this kind of contribution carries far more weight than a flashy demo video.

In robotics, the open-source movement is experiencing a critical period of acceleration. From ROS (Robot Operating System) laying the software infrastructure groundwork, to the wave of research democratization sparked by the open-sourcing of the MuJoCo physics engine, to the emergence of integrated hardware-software open-source platforms like Microduck — each step lowers the barrier to entry and expands the community of participants. History has shown time and again that when more brilliant minds can participate in the same technical direction at low cost, the speed and diversity of innovation grow exponentially.

As the author hoped: may people "have fun, learn a lot, or both" on this platform.