Sudu Technology Debuts Advanced Robotics at WAIC, Showcasing Simulation-to-Real-World Prowess

At the recent World Artificial Intelligence Conference (WAIC), Sudu Technology, a rapidly emerging embodied intelligence firm, made a significant domestic debut, showcasing its advanced robotics capabilities. Founded just over a year ago and already valued at RMB 20 billion, Sudu presented four sophisticated demonstrations that aim to answer a fundamental question in the field: can models trained purely in simulation effectively operate in the real world? The company's CEO, Han Zheng, confidently stated that their approach achieves over 99% success rates, a critical benchmark for commercial viability.

Four Demos, One Core Message

Sudu's presence at WAIC featured four distinct demonstrations, each designed to push the boundaries of embodied AI. These included precision assembly requiring sub-millimeter accuracy with dual-hand coordination, flexible packaging involving intricate manipulation of soft materials, mobile manipulation for end-to-end closed-loop control in dynamic environments, and a battery module production line developed in collaboration with CATL, showcasing four-robot synergy. These demos collectively illustrate Sudu's commitment to bridging the gap between simulation and real-world application, moving beyond simpler pick-and-place tasks to more complex, contact-rich operations.

Tackling Extremes: Precision and Flexibility

Responding to industry skepticism about the applicability of simulation in contact-rich scenarios, Sudu presented two challenging demos: precision assembly and flexible packaging. The precision assembly task, involving the sub-millimeter accurate fitting of components, tested the limits of force control and pose estimation, requiring real-time adjustments based on tactile feedback. Conversely, the flexible packaging demo highlighted the difficulty of modeling and manipulating deformable objects like fabric, where each interaction drastically alters the state. Sudu demonstrated that a single base model, trained through simulation, could successfully handle both these extreme tasks, maintaining a high level of anti-interference capability comparable to their earlier picking successes.

Bridging the Gap: Mobile Manipulation and Multi-Robot Collaboration

Beyond static manipulation, Sudu also addressed the complexities of mobile manipulation. Unlike many existing systems that rely on extensive real-world data collection, which is difficult for mobile platforms, Sudu employs an end-to-end whole-body control strategy. This approach integrates perception, locomotion, and manipulation into a unified policy, enabling robots to perform tasks like opening doors or loading materials in dynamic environments. The most ambitious demonstration was the battery module production line, a collaborative effort with CATL. This showcased four robots working in concert to replicate a real industrial process, from material loading to final assembly. This multi-robot coordination highlights Sudu's ability to manage state transfer and task sequencing in long-horizon operations, a critical step towards complex industrial automation.

The Simulation-First Philosophy

Sudu's core technical strategy revolves around a simulation-first approach, leveraging large-scale virtual data to build a robust foundation model. CEO Han Zheng explained that while real-world data is crucial for fine-tuning, it is not scalable for initial training due to cost and time constraints. Sudu's method involves pre-training a highly generalizable base model in simulation and then fine-tuning it with a small amount of real-world data for specific deployment. This hybrid approach, termed Sim2Real, aims to balance data efficiency with generalization accuracy, a path Han Zheng believes is the most scalable for commercial deployment.

A Novel World Model for Embodied AI

Central to Sudu's strategy is their unique interpretation of a world model. While existing models focus on rendering, simulation, or high-level planning, Sudu proposes a fourth category: one that achieves extreme granularity in rendering and simulation. This involves a high-precision geometric understanding and prediction of the environment and objects, including material properties, friction coefficients, and dynamic constraints. This detailed physical fidelity is essential for contact-rich manipulation tasks, enabling robots to learn transferable physical laws through extensive virtual practice. This integrated approach, combining world models with reinforcement learning, allows robots to learn and evolve within a physically accurate virtual environment.

Long-Term Vision and Hardware Independence

Sudu's commitment to simulation is part of a broader long-term vision, prioritizing foundational model development over creating flashy demos. While the company has developed its own hardware, including robotic bodies, this was a strategic decision driven by the lack of suitable off-the-shelf components for their advanced algorithms. They envision a future akin to the "iPhone + iOS" model, where their hardware and foundational models serve as a platform for third-party application development. This approach allows them to maintain control over the critical hardware-software integration needed for state-of-the-art embodied AI, while also fostering an ecosystem.

The 99%+ Reliability Imperative

CEO Han Zheng repeatedly emphasized that 99%+ success rate is the non-negotiable prerequisite for commercialization in embodied AI. He contrasted this with the challenges faced by autonomous driving, where multi-agent interaction at high speeds is paramount. For robots, the focus is on high-fidelity physical interaction and precise control. Sudu's simulation-centric approach, coupled with their advanced world model and hardware integration, is designed to meet this stringent reliability requirement, enabling robots to tackle complex, real-world tasks reliably and scalably.