BRIDGE: An Open-Source Humanoid Platform via Morphology-Control Co-Design for Physical AI

Conventional humanoid development follows a decoupled paradigm: hardware design and whole-body control are engineered separately, and human behavioral data is treated as an afterthought for a body that was never shaped for it. The result is systems that compromise the fluidity and agility of human-like movement, which limits the vision of general-purpose embodiment in physical AI. BRIDGE starts from the opposite direction — let data about how humans move drive both what the robot is and how it is controlled.

Core Innovation

• Data-driven morphology-control co-design — a framework that jointly optimizes humanoid morphology and control for human-like movement, replacing the decoupled hardware-then-control pipeline.
• A joint fidelity metric — a new evaluation quantity that simultaneously considers kinematic retargeting fidelity to human motion and dynamic tracking performance, quantifying how faithfully a morphology can reproduce human behavior.
• A complete open-source release — the framework is realized in Bridge, an 88 cm-tall humanoid whose design, control policy, videos, and open-source materials are made public.

Results

• State-of-the-art performance across all reported metrics compared with baseline humanoids including Bumi, K1, and Toddlerbot.
• Superior fidelity in capturing human motion data on the real platform.
• Demonstrated capability across foundational locomotion, robust balance, and highly dynamic maneuvers.

Limitations

Bridge is a compact 88 cm platform, so the abstract’s evidence does not yet speak to full-size humanoid regimes or manipulation-rich tasks. The comparison set is a handful of existing small humanoids, and the headline results are morphological-fidelity and control metrics rather than long-duration deployment studies. As with any data-driven morphology result, the framework’s output quality depends on the breadth of the human-motion corpus used.

Industry Implications

For the humanoid industry, BRIDGE matters in two ways. First, it is an open, reproducible small-scale platform with a working whole-body policy — a practical starting point for teams validating control stacks, sim-to-real pipelines, and data collection before committing to full-size hardware. Second, the morphology-control co-design methodology offers a concrete route toward bodies that are literally shaped to imitate human data well, which is the core value proposition of general-purpose humanoid embodiment. Open sourcing both the robot and the policy lowers the entry barrier for physical-AI research and product exploration alike.