Pudu Robotics Launches PuduFM 1.0 Foundation Model and PuduAgent Platform for Embodied AI

Pudu Robotics has launched two foundational technologies for embodied intelligence: PuduFM 1.0, a physical-intuition-driven robot foundation model, and PuduAgent, a general-purpose embodied agent software platform. The releases support Pudu's structural strategy of 'One Brain, Multiple Embodiments,' allowing diverse robotic systems—ranging from delivery and cleaning units to industrial AMRs and humanoid systems—to share a single, end-to-end cognitive framework.

PuduFM 1.0 features a dedicated Physical Intuition Model (PIM) and a deeply aligned Vision-Language-Action (VLA) model. By transitioning from traditional pattern matching to predictive latent-space dynamics, the system can foresee physical interactions and contact constraints in real time. Rather than relying on pixel-level visual forecasting, PIM utilizes sparse state predictions to maximize computational efficiency during complex real-time manipulation and navigation tasks.

To bridge the gap between cognitive modeling and physical execution, the PuduAgent platform introduces a structured system, capability and safety layer architecture. The system layer manages multi-stage task decomposition and includes a three-tier memory setup (working, episodic and compressed memory) to sustain goal consistency across long-horizon operations. This configuration helps eliminate task drift during workflows that exceed 30 minutes.

Through its open capability layer, PuduAgent abstracts atomic skills like contour navigation, obstacle avoidance and elevator interaction into reusable modules. This allows different robot forms to access accumulated environmental maps and operational intelligence without custom programming. The platform also integrates an ecosystem featuring a complete SDK, simulation environments and the SkillHub marketplace for third-party application developers.

The architecture is actively deployed on Pudu's full-category product matrix, including the newly introduced PUDU D7 industrial semi-humanoid robot, the BellaBot series, and the D9 full-size humanoid robot. Backed by a global commercial market share of 23 percent and over 130,000 units shipped, Pudu aims to utilize its field deployment scale to create a self-reinforcing data flywheel for continuous model refinement.