Yuandian Robotics and CMU Launch Full Open-Source Ecosystem for Humanoid Robots

The humanoid robot industry has seen rapid hardware progress, with machines capable of backflips and dynamic martial-arts moves, but the software ecosystem remains fragmented. The report noted that Unitree's G1 robot achieved a 96.36 percent backflip success rate and a 10-millisecond inference delay with the OmniXtreme framework. Yet different manufacturers use different body configurations, joint setups, sensor interfaces and calibration methods, so algorithms and training data developed for one robot often cannot transfer to another. Yuandian Robotics is seeking to address this problem by building a common open-source platform.

The first paper, called BRIDGE, offers an open-source, AI-native hardware design methodology. Instead of starting from a mechanical design and then trying to make algorithms work on it, the team uses human motion data as a starting point and optimizes the robot's body to match that data. Using the SMPL parametric human model, they derived an 88-centimeter research platform with 21 active joints through a four-step workflow that includes motion compression, physics-aware instantiation, dynamic verification and failure-driven self-improvement.

To decide which waist degrees of freedom to keep, the team first compared kinematic re-targeting errors, with a combination of roll and yaw giving the lowest error of 0.02570 among three candidates. After dynamic simulation, keeping only yaw produced a tracking error of 0.02115, lower than keeping only roll at 0.02311. This led to a counterintuitive conclusion: for an 88-centimeter robot, waist yaw is more beneficial for dynamic stability than pitch.

BRIDGE's hardware is said to cost about $1,500, around a quarter of the estimated $6,000 for an alternative platform called Toddy. In a quantitative human-likeness test, BRIDGE scored 0.5252 against Toddy's 0.3883, and its body-proportion scaling factor was 1.021, much closer to the SMPL reference than Toddy's 1.171 or other platforms' values above 1.3. The design also kept its center of mass within the support polygon while imitating actions such as touching toes, standing on one leg and making low-center turns. The platform is fully open: code, design files and a co-design interface are all provided.

The second paper, FlowMimic, proposes an open-source control framework for handling heterogeneous instructions. Commands can come from VR controllers, motion-capture suits or AI motion generators that have different data formats and noise levels. Rather than first filling in a complete pose and then tracking it, FlowMimic lets the control policy directly learn to deduce missing information from sparse input.

Yuandian Robotics also formally launched the developer-facing OpenBridge ecosystem, which is intended to combine the hardware and control toolchains into a comprehensive open ecosystem. The company said the effort represents a path from closed development to open collaboration in robotics research and development, giving developers a common platform to modify and extend the system.