Gaode’s World Model Enables Real‑Time Physical Interaction and Scene Construction

During the 2026 Global Digital Economy Conference, Alibaba Gaode presented “Gaode Space Intelligence: Practical Applications of World Models in Physical Interaction and Scene Construction,” introducing the ABot full‑stack embodied AI system, the autonomous robot Gaode Tutu, and the real‑time interactive world model DreamX‑World.

ABot is built on a three‑layer infrastructure: the physical‑world layer (ABot‑World, ABot‑Earth) handles world generation and evolution; the physical‑action layer (ABot‑N, ABot‑M) provides a “one brain, multiple bodies” framework for embodied navigation and manipulation; the physical‑agent layer (ABot‑ER, AgenticOS) delivers self‑evolving model, memory, skill, and harness capabilities. ABot‑World Studio can create an interactive world with a single RTX 5090 GPU; ABot‑Earth covers more than 190 countries, generating city‑scale 3D maps in minutes; ABot‑N0 completed autonomous navigation of nearly 3 km in Beijing’s evening rush and won multiple SOTA benchmarks such as WorldScore, WorldArena, AgiBot, and Embodied Arena.

Gaode Tutu addresses the shortage of guide dogs for China’s 17 million visually impaired people. It features a central AI compute unit, all‑weather perception, a 60 kg static load capacity, four‑hour battery life, and a detachable guide‑dog kit. The robot supports point‑to‑target, object‑search, follow, and zero‑hand‑over long‑range tasks, and includes Map‑as‑Memory and multimodal memory functions. In a live demo, Tutu achieved zero‑hand‑over navigation over a 5 km stretch of Beijing’s CBD during peak hours, validating the L1 world layer, L2 base layer, and L3 scheduling layer that enable a Language→Spatial→Physical evolution.

DreamX‑World is characterized by long‑term consistency, real‑time interactivity, and consumer‑grade GPU inference. It shifts AI from merely describing the world to predicting how the world changes when conditions are altered. The model tackles three core challenges: maintaining long‑term temporal consistency, delivering sub‑second interactive latency, and reducing inference cost. It achieves one‑minute long‑term consistent video generation, sub‑second response time, and 16 FPS real‑time inference on eight NVIDIA RTX 5090 GPUs (≈5090 fps = 16). Benchmarks VBench and Omni‑WorldBench both report state‑of‑the‑art performance.

The technical stack follows a data‑training‑inference triad. The data side incorporates Unreal Engine‑generated first‑person and third‑person samples with precise camera and action control, followed by extensive cleaning and annotation. Training introduces the efficient camera control method E‑PRoPE, reducing inference cost by about 30 %; long‑term memory retrieval keeps “Rotate‑Away‑Rotate‑Back” scenes consistent; a pipeline combining bidirectional‑to‑unidirectional conversion, Causal Forcing, DMD Distillation, and a 60‑second rollout yields one‑minute generation; structured events are merged into text conditions for natural world‑change understanding; reinforcement learning supplements diversity and motion quality after few‑step distillation. Inference employs DMD few‑step generation, mixed precision, sequence parallelism, operator fusion, feature reuse, and VAE pruning, positioning the model at the forefront of open‑source performance.

Overall, the session mapped Gaode’s two‑decade spatial‑data foundation to a three‑layer embodied AI stack and demonstrated concrete products—from the guide‑dog robot to the interactive world model—forming a “spatial intelligence map” that bridges digital maps, physical AGI, and consumer/industry applications.