Home / News

Chengdu Unveils 5th-Gen Gongga Humanoid: Spatiotemporal Memory, Causal Planning and a World Model

Date:11-19-2026

The Chengdu Humanoid Robot Innovation Center launched the fifth-generation Gongga humanoid robot on 11 November 2026, framing the release around its "embodied brain" rather than new hardware. The center claims the system represents the front rank of humanoid robot cognition and positions it for unstructured home environments.

Gongga has iterated unusually fast: the first generation appeared in December 2024, followed by March 2025, August 2025, April 2026 and now November 2026. Five generations in 23 months indicates a software-led development loop with a stabilised mechanical platform.

What changed in generation five

  • Structured perception: RGB, depth and proprioception fused into a hierarchical 3D scene graph, updated incrementally as objects move
  • Spatiotemporal memory engine: stores spatial coordinates, object physical state, temporal evolution and past interaction outcomes, so the robot can recall where an item was, when it moved and whether a prior attempt worked
  • Causal planning: task decomposition based on causal dependencies and preconditions rather than statistical correlation, with replanning when the environment changes
  • Latent action-conditioned world model: multi-step prediction and candidate-action simulation before execution, plus closed-loop trajectory correction

Why it matters

Most humanoid demonstrations are single-episode: the robot is placed in a prepared scene, performs a rehearsed task, and forgets. The limiting factor for home deployment is not grasping capability but continuity — knowing that the cup was moved to the dishwasher yesterday, and that the last attempt to carry two glasses failed. Memory and world-model prediction are the two mechanisms the industry has converged on to address this, and Gongga Gen 5 is one of the clearest public statements of that architecture from a Chinese developer.

The choice of home environments as the target is also a strategic bet. Homes are the least structured, lowest fault-tolerance setting, which makes them a poor first market commercially but a strong forcing function technically.

What is not yet proven

The center has released capability descriptions and demonstration video, not benchmark results. Task success rates over multi-day horizons, recovery behaviour after failure, and the compute payload required to run the world model on-board are all undisclosed. No price, production volume or customer has been announced.

Source

Reported by Science and Technology Daily, People's Daily Sichuan and Sichuan Online, 11-12 November 2026. Chinese-language sources only.

Prev: China

Next: AMEC Ships First Primo Matrix Plasma Wafer Dicing Tool as Advanced Packaging Demand Climbs