Pudu Robotics

One-brain / multi-form architecture: Pudu repeatedly frames '一脑多形' (one brain, many shapes) and '具身智能' (embodied intelligence) as core. Technically this implies a shared, high-level policy/representation layer that can be retargeted to very different morphologies (floor scrubbers, humanoid-like greeters, tuggers). That is more than marketing—it's an architectural choice that centers on transferable task representations and morphology-specific low-level controllers rather than rebuilding stacks per product.

Vertical hardware-software integration down to joint modules and motion controllers: they claim full-stack self-research (navigation algorithms, multi-robot scheduling, robot motion controllers, integrated joint modules). Many service-robot players outsource actuators or controllers; Pudu is building custom motor drivers, joint modules and real-time control stacks. This enables torque-level control, tighter feedback loops, and dynamic behaviors that commodity controller + ROS setups struggle to match.

AI-native large cleaning robot (BG1) — a productized ML-first approach to a safety- and reliability-critical domain: describing BG1 as 'AI原生' suggests learned perception-to-action pipelines for surface classification, adaptive scrub patterns, and heterogeneous-floor handling (instead of purely geometric SLAM + hand-tuned heuristics). That requires sensor fusion, online state estimation and adaptive force/pressure control tied to low-level motor controllers.

Fleet-scale multi-robot scheduling + heterogeneous fleet management: they emphasize multi-robot scheduling and global service network. Solving this across cleaning, delivery and concierge robots implies real-time task allocation, cross-robot collision avoidance in mixed-use human environments, charging/station scheduling, and potentially global policy updates—non-trivial distributed systems challenges beyond single-robot autonomy.

Data and deployment-driven sim-to-real and continual learning pipeline (implied): claiming 120k+ cumulative units and a global footprint suggests they can collect diverse operational data. The interesting technical implication is a continuous loop: field data → simulation / domain-randomized training → per-site adaptation and OTA model updates. Handling dataset drift, safety/regression testing, and fast rollout are hidden complexities.