Article · AI translation
IT Home reported on November 11 that the Chengdu Humanoid Robot Innovation Center unveiled the fifth-generation Gongga humanoid robot today, claiming it has “the world’s most advanced core technology for humanoid robot brains.”
Described as the culmination of the Chengdu Humanoid Robot Innovation Center’s work on embodied-intelligence brain technology, the fifth-generation Gongga is equipped with a self-developed integrated embodied-intelligence system. Its core modules, including scene perception, spatiotemporal memory, causal reasoning and planning, and an action world model, have undergone systematic upgrades. The center claims it “represents the highest level of humanoid robot brain technology in the world’s current leading tier.”
The fifth-generation Gongga’s embodied brain integrates the complete technical chain of “structured perception — spatiotemporal memory — causal planning — action simulation — execution feedback,” enabling end-to-end coordination of environmental understanding, memory updates, task planning, and action execution.
For environmental perception, the robot combines RGB vision, depth information, and data on its own pose to build a hierarchical three-dimensional scene graph. Using cross-view entity association, semantic relationship modeling, and incremental scene updates, it continuously creates a structured representation of a changing home environment. It can reliably identify the locations, categories, and arrangement of objects, and adapt to frequent changes in homes, such as objects being moved or new clutter appearing.
For memory, the innovation center’s self-developed physical spatiotemporal memory engine stores and manages spatial coordinates, objects’ physical states, patterns of change over time, and past interaction experience in a unified system. The robot can maintain consistent object identities over long periods and supports retrieval of sequential memories, alignment of memories across multiple rounds, and automatic resolution of conflicts. Put simply, it can “remember” where things were previously placed, when they were moved, and whether past actions succeeded or failed. This addresses the lack of memory across individual tasks in traditional robots and supports long-running household tasks spanning multiple days.
The decision-making and planning module introduces a causal reasoning framework, so it no longer selects actions based solely on statistical associations in data. The system automatically identifies causal dependencies within a task, prerequisites for actions, expected outcomes, and constraints on the robot’s own resources, and independently breaks tasks down into a hierarchy. When the environment changes unexpectedly or a task is obstructed, it can dynamically replan in real time, assess unexpected situations, and adjust its approach on its own instead of simply stopping the task.
A key feature of the fifth-generation Gongga is its action execution layer, which uses a latent-space action-conditioned world model. The robot can predict future states multiple steps ahead, simulate candidate actions, and assess uncertainty and risks, first anticipating the outcomes of different actions in a virtual space. It then combines motion-trajectory optimization with closed-loop predictive control to correct errors in its actions in real time, substantially increasing the success rate of delicate household tasks such as grasping, putting away, and wiping, while reducing mistakes such as collisions and dropped objects.
IT Home note: The Chengdu Humanoid Robot Innovation Center’s full name is Chengdu Humanoid Robot Innovation Center Co., Ltd. Established on April 11, 2024, it is the first originating project under Sichuan Province’s “No. 1 Innovation Project” and the first humanoid robot innovation center in central and western China.