AMAP Spatial Intelligence Models & Systems
Understand · Predict · Simulate · Create · Plan · Act in the real world.
DreamX is the spatial intelligence portfolio of Alibaba AMAP, built and released by the DreamX team through AMAP-ML.
We connect research, production systems, and real-world deployment across maps, mobility, local services, digital content, and interactive worlds.
The team releases open-source projects, benchmarks, and publications at ICLR, CVPR, ECCV, ACL, AAAI, SIGGRAPH, SIGGRAPH Asia, IJCV, ICCV, ICML, KDD, CIKM, EMNLP, ACM MM, and WWW.
Flagship releases · Research directions · Recent updates · All projects · Collaborate with us
40+ open-source projects · 52 papers at major venues in 2026 · 3 core problems · 6 DreamX model and system families
| Understand & Predict | Generate & Simulate | Plan & Act |
|---|---|---|
| Maps, mobility, places, intent, and urban dynamics | World models, media, scenes, and digital assets | Agents, tool use, decision-making, and embodied systems |
| DreamX-Predictor · DreamX-REC | DreamX-World · DreamX-Creator | DreamX-Agent · DreamX-Phi |
| Project | What it brings to spatial intelligence | Links |
|---|---|---|
| DreamX-World 1.0 | Interactive, long-horizon world simulation with an open 5B model. | Code · Report |
| DreamX-Phi 1.0 | Geometry-aware, action-conditioned video world modeling for robotic manipulation. | Code · Paper |
| LongHorizon-Harness | Durable task state and independent verification for reliable computer-use agents. | Code · Paper |
| MobilityBench | Real-world benchmark for route-planning agents in AMAP-native mobility scenarios. | Code |
| SkillClaw | Evolves reusable agent skills from real interaction traces. | Code |
| FluxText | Controllable scene-text editing for practical visual-asset creation. | Code |
Browse the complete project map →
We build AI that understands the real world and its change over space and time, predicts what comes next, generates and simulates digital counterparts, plans toward human goals, and acts through products and embodied systems.
For AMAP, this closes a learning-and-deployment loop across maps, routes, places, mobility, urban environments, local services, digital content, and physical action.
| Direction | DreamX families | Representative work |
|---|---|---|
| Understand and predict the world | DreamX-Predictor · DreamX-REC | MobilityBench, Thinking-with-Map, SocioReasoner |
| Generate and simulate the world | DreamX-World · DreamX-Creator | DreamX-World, OmniDance, RL3DEdit |
| Plan and act in the world | DreamX-Agent · DreamX-Phi | DreamX-Phi, AutoDrive-R2, GPG |
All DreamX families share spatial data and knowledge, multimodal foundation models, spatiotemporal and world modeling, agent and reinforcement-learning methods, generative modeling, plus scalable infrastructure and evaluation.
| Family | Role |
|---|---|
| DreamX-Predictor | Models the evolution of traffic, mobility, demand, supply, and urban conditions. |
| DreamX-World | Learns controllable, persistent, physically grounded, interactive world models. |
| DreamX-Agent | Reasons over spatial context, uses tools, and completes map, mobility, and local-service workflows. |
| DreamX-Phi | Connects perception, reasoning, decision-making, and physical action. |
| DreamX-REC | Matches intent with places, content, routes, and services under real-world constraints. |
| DreamX-Creator | Generates and edits map assets, local-service content, images, videos, 3D scenes, and spatial media. |
- 2026.08.21 — The DreamX team has 3 papers accepted to EMNLP 2026, advancing multimodal continual learning, self-evolving agents, and language-model reasoning.
- 2026.08.13 — DreamX-Phi 1.0 launches a geometry-aware, action-conditioned video world model for robotic manipulation, leading WorldArena 2.0 Track 1 in the August 12 leaderboard snapshot. Paper
- 2026.08.03 — LongHorizon-Harness introduces a Manage–Execute–Audit loop for durable, verified computer-use progress. Paper
- 2026.07.20 — OmniDance is selected as an ECCV 2026 Oral, advancing large-scale multimodal dance-video generation from text, image, and music.
- 2026.06.18 — Five DreamX papers are accepted to ECCV 2026, spanning spatial intelligence, generative modeling, and multimodal AI.
- 2026.06.15 — DreamX-World releases its 1.0 technical report and the DreamX-World-5B model for one-minute interactive world generation. Report
Earlier updates
- 2026.07.20 — The DreamX team adds four publications across SCALAR++ (IJCV), Evaluation-Verification Reward (SIGGRAPH Asia 2026), MAR-GRPO, and Peak-End-Net (ACM MM 2026).
- 2026.05.18 — MobilityBench is accepted to KDD 2026 as an Oral paper.
- 2026.05.12 — CoEvolve and Thinking-with-Map release work on agent evolution and map-augmented reasoning (ACL 2026).
- 2026.05.01 — UniMRG, Train-Free Infinite-Frame Generation, D-Evo, and EEPO are accepted to ICML 2026.
- 2026.04.22 — DCW and EMF are accepted to CVPR 2026.
- 2026.04.01 — MACE-Dance releases a music-driven dance-video generation framework (SIGGRAPH 2026).
- 2026.03.23 — Omni-WorldBench, AutoDrive-R2, Video-STAR, and RL3DEdit release work across world modeling, driving, video understanding, and 3D editing.
Understand and predict the world
| 仓库 | Contribution | Venue |
|---|---|---|
| MobilityBench | Route-planning agent evaluation in real-world mobility scenarios. | KDD 2026 Oral |
| Thinking-with-Map | Map-augmented geolocalization agent trained with reinforcement learning. | ACL 2026 Findings |
| SocioReasoner | Vision-language reasoning for urban socio-semantic segmentation. | ICLR 2026 |
| DSFNet | Multi-scenario route ranking with a public industrial driving-route dataset and AMAP deployment. | WWW 2025 |
| GenMRP | Generative multi-route planning for efficient, personalized, real-time industrial navigation. | CIKM 2026 |
| IntSR | Integrated generative modeling for search and recommendation across AMAP scenarios. | CIKM 2026 |
| IntTravel | Real-world dataset and generative framework for integrated multi-task travel recommendation. | arXiv 2026 |
| FE2E | Image-editing priors for dense geometry estimation. | CVPR 2026 |
| UniVG-R1 | Reasoning-guided universal visual grounding with reinforcement learning. | arXiv 2025 |
| Taming-Hallucinations | Counterfactual video generation for reducing MLLM video hallucinations. | — |
Generate and simulate the world
| 仓库 | Contribution | Venue |
|---|---|---|
| DreamX-World 1.0 | General-purpose world model for interactive world simulation. | — |
| Code2World | GUI world model via renderable code generation. | — |
| FluxText | Diffusion transformer baseline for scene-text editing. | — |
| OmniDance | Multimodal dance-video generation from text, image, and music. | ECCV 2026 Oral |
| SCALAR / SCALAR++ | Efficient controllable generation through scale-wise visual autoregressive learning. | AAAI 2026 / IJCV |
| MAR-GRPO | Stabilized RL for autoregressive-diffusion image generation. | ACM MM 2026 |
| RL3DEdit | Geometry-guided RL for multi-view-consistent 3D scene editing. | ECCV 2026 |
| MACE-Dance | Motion–appearance cascaded generation for music-driven dance video. | SIGGRAPH 2026 |
| Omni-Effects | Prompt-guided, spatially controllable composite visual-effects generation. | AAAI 2026 |
| S2-Guidance | Training-free stochastic self-guidance for diffusion models. | ICLR 2026 |
| EPG | Pixel-space generative modeling via self-supervised pre-training. | ICLR 2026 |
| USP | Unified self-supervised pretraining in VAE space for diffusion models. | ICCV 2025 |
| EMF | Text-conditioned one-step image generation. | CVPR 2026 |
| DCW | Differential correction for SNR-t bias in diffusion probabilistic models. | CVPR 2026 |
| NarrLV | Narrative-centric evaluation for long-video generation models. | ICLR 2026 |
| Imagery搜索 | Adaptive test-time search for video generation. | AAAI 2026 |
| Eevee | High-resolution benchmark for video-based virtual try-on. | CVPR 2026 Findings |
| VMBench | Perception-aligned benchmark for video motion generation. | ICCV 2025 |
Plan and act in the world
| 仓库 | Contribution | Venue |
|---|---|---|
| DreamX-Phi 1.0 | Geometry-aware, action-conditioned video world modeling for bimanual robotic manipulation. | arXiv 2026 |
| LongHorizon-Harness | Verified long-horizon computer use through durable task state and Manage–Execute–Audit loops. | arXiv 2026 |
| SkillClaw | Agentic evolver for collective skill-library improvement. | — |
| AutoDrive-R2 | Reasoning and self-reflection for VLA models in autonomous driving. | ICLR 2026 |
| Tree-GRPO | Tree-search rollouts for LLM-agent reinforcement learning. | ICLR 2026 |
| GPG | Simple and strong group policy-gradient baseline for model reasoning. | ICLR 2026 |
| CoEvolve | Agent–data mutual evolution for training LLM agents. | ACL 2026 |
| Role-Agent | Dual-role evolution that trains an LLM as both an agent and an environment model. | EMNLP 2026 |
| MathForge | Difficulty-aware GRPO and multi-aspect reformulation for math reasoning. | ICLR 2026 |
| Video-STAR | Tool-using RL for open-vocabulary action recognition. | ICLR 2026 |
Shared foundations and evaluation
| 仓库 | Contribution | Venue |
|---|---|---|
| SpatialGenEval | Spatial-intelligence evaluation for text-to-image models. | ICLR 2026 |
| Omni-WorldBench | Benchmark for interactive response capabilities of world models. | arXiv 2026 |
| RealQA | Realistic image-quality and aesthetic scoring with multimodal LLMs. | — |
| M2Note | Training-free continual evolution for vision-language models through an editable mistake notebook. | EMNLP 2026 |
| Peak-End-Net | Peak-end-rule-inspired framework for generalizable video aesthetic assessment. | ACM MM 2026 |
| Evaluation-Verification Reward | Fine-grained reward modeling for consistent multi-reference image editing. | SIGGRAPH Asia 2026 |
| FASA | Frequency-aware sparse attention for efficient sparse decoding. | ICLR 2026 |
We welcome research interns, full-time researchers, AI engineers, and academic collaborators working on spatial intelligence, LLM agents, reinforcement learning, world models, multimodal learning, embodied AI, recommendation, and generative AI.
Please send your CV, representative work, and research interests to cxxgtxy@gmail.com. You can also visit our organization page or team homepage.