Daxiao Robotics Deploys Dual-Brain Architecture for Autonomous Hotel Laundry Services

Oriental Daily reporter Cao Lei reported on November 10: From taking clothes from the guest room, walking through the corridor, operating the washing machine and dryer, to putting the dried clothes in a basket and returning them to the guest room, a complete hotel laundry service process is being autonomously completed by robots. Daxiao Robot will implement the dual-brain collaborative architecture in the hotel laundry scene for the first time, relying on the Kaiwu world model, general model, and Embodied Intelligence RSI strategy, building a capability chain of “digital pre-training—real machine deployment—self-iteration,” marking the robot's transition from single action execution toward full-process continuous service.

This implementation focuses on demonstrating the synergistic value of the Mundo world model, the GPT-Astra general model, and the RSI recursive self-improvement strategy. Among them, Mundo, as an important supplement to the Enlightenment world model system, can reconstruct high-fidelity digital twin scenes based on real hotel images, not only restoring spatial structures and positional relationships but also incorporating physical attributes such as object friction and mass, utilizing a physics engine to simulate collisions and interactions. Relying on this explorable, editable, and interactive virtual training space, robots can pre-rehearse a series of complex actions such as navigation, obstacle avoidance, opening doors and placing clothes, and pressing buttons to start, repeatedly optimizing strategies through the Real2Sim2Real training chain, significantly reducing on-site adaptation costs.

If world models allow robots to “understand the environment,” general models represented by Astra are responsible for “understanding tasks and generating skills.” Hotel laundry involves long-process planning, flexible object grasping, and precise equipment operation; relying solely on real-time inference from general models would result in high latency. Daxiao Robot's solution is to let Astra handle task decomposition, plan exploration, and skill generation, then solidify the validated action logic into Skills that can be called at high speeds. Test data shows that when unoptimized, a single new operation takes over 30 minutes, but after being deposited as a solidified skill, one round of execution takes only about 18.648 seconds, a speedup of approximately 97 times. The accompanying developed BotAgent can also organize 53 atomic actions into 14 high-level steps, flexibly combining over 200 types of scene actions to support complete laundry order scheduling.

The core of continuous evolution is the RSI recursive self-improvement strategy designed specifically for embodied intelligence. Robots record images, motion trajectories, and feedback data during real-machine execution, and Astra identifies the reasons for failure and outputs optimization plans; the modifications are first verified through offline playback, and only after confirming effectiveness are they deployed to the real machine for iteration. This system also builds in autonomous resetting, independent evaluation, and version rollback mechanisms, allowing robots to continuously expand their skill boundaries through repeated trial and error. Taking the removal of clothes from a dryer as an example, after 165 rounds of real-machine experiments and approximately 80 version iterations, the proportion of jeans remaining in the drum significantly decreased, fully proving that real physical feedback is the key driving force for upgrading robot capabilities.

For a long time, world models and general intelligence have mostly remained at the simulation and laboratory stages. This implementation in the hotel laundry scene is a concentrated verification of the dual-brain collaborative architecture in a real service scenario: the world model provides a virtual training ground, the general model is responsible for skill creation, and the RSI strategy continuously polishes the implementation effects. When scenes can be reconstructed, actions can be pre-simulated, and strategies can self-iterate, robots are no longer just executors of preset programs, but possess the potential to adapt to complex real environments and continuously optimize service capabilities. This paradigm also provides a reference for embodied intelligence to move toward more industrial scenes, allowing intelligent robots to truly enter daily service spaces such as hotels and supermarkets.