Skip to content
Amap Phys AI Data: Spatial Data Platform for Physical AI Training, Simulation & Deployment

Amap Phys AI Data: Spatial Data Platform for Physical AI Training, Simulation & Deployment

Amap unveiled Phys AI Data at the 2026 World Internet Conference, a three-layer spatial data foundation comprising training data (embodied, spatial, synthetic), a simulation evaluation platform, and Physical AI Map, enabling robot training, validation, and deployment in real-world scenarios like malls and factories.

Amap Tech
Amap Tech
Amap Tech

Background: From Mapping for Humans to Mapping for Machines

At the 2026 World Internet Conference in Wuzhen, Amap solution product expert Wang Miyang presented Phys AI Data , a one-stop spatial data foundation for physical AI training and application. For two decades Amap mapped the world for people; now it maps the world for machines — converting real-world space, behavior, and interaction into data that models can learn from, verify, and deploy.

Three-Layer Architecture: Training Data, Simulation Evaluation, Physical AI Map

The platform is organized into three layers, each addressing a distinct need: robots must know how to act, understand their space, and be validated before deployment.

1. Training Data — Three Categories

Embodied Data (Ego & UMI): Ego data is collected in the wild — not staged in labs — across real business scenes such as precision assembly, cosmetic care, and cooking. UMI gripper data matches the target real robot, with pose reproduction error ≤5 mm and an 80% deployment success rate under corresponding tasks and test conditions. The dataset covers 10+ robot morphologies for cross-morphology re-collection, or can be customized per customer-specified scene, task, and robot body.

Spatial Base Data: Comprises real-scene asset collection (street-view, aerial, POI panoramic video, road panoramic) covering air, streets, and indoors, plus 3D reconstruction: point clouds for geometry and 3D Gaussian Splatting (3DGS) for appearance. Acquisition follows three professional tiers: professional scanning, panoramic cameras, and image-based generation.

Synthetic Data: Hundreds of thousands of hours of stock data with frame-aligned RGB and depth, geometry and appearance paired on the same trajectory, and replaceable weather and characters — suitable for world-model training.

2. Simulation Evaluation — Private-Deployable Platform & Scene Assets

Offers a simulation evaluation platform that can be deployed privately, with over 100,000 high-precision assets supporting physics effects (friction, fluid), weather, lighting, and scene configuration. Simulation scenes are built from real spaces and output multimodal data: multi-camera, depth, segmentation, optical flow, and pose. The boundary is clear: the goal is to connect real scenes, simulation verification, and real-robot cross-check — not to equate simulation scores directly with real-world capability.

3. Physical AI Map — Topology, Localization Anchors, Spatial Semantics

Provides topological road networks, localization anchors, and 200+ spatial semantic categories. Coverage exceeds 90% of high-traffic malls nationwide, with indoor positioning accuracy better than 1 meter. Already applied in large venues and robot-dog visual-language navigation (VLN) scenarios.

Training Chain: Human Video → Simulation → Real Robot, Driven by a Data Evolution Flywheel

The training pipeline relays three data layers:

Human operation videos captured via head-mounted (Ego) and hand-mounted (UMI) devices — low cost, large scale, giving models a basic grasp of manipulation.

Simulation data generated in virtual environments — enables repeated trial-and-error in extreme or robot-damaging scenarios, with massive parallel training.

Real-robot data from physical environments — higher quality and cost, used for final fine-tuning and precision calibration.

A data evolution flywheel connects the layers: failure cases drive targeted re-collection and data synthesis; evaluation results feed back into the next training round; data and models iterate synchronously.

Applications & Ongoing Work

Phys AI Data is already used to train Amap's Tutu embodied products, covering tasks such as target understanding, path planning, and on-site traffic adaptation for guiding, inspection, and delivery.

The ABot system model trained on this foundation has been integrated into humanoid robots, with R&D focusing on home, retail, and industrial scenes — progressing from flexibility to dynamics to precision.

Physical AI requires not only model innovation but also stable data supply, verifiable evaluation, and real deployment scenarios. Amap continues to refine the physical AI data foundation to bring robots into more real-world settings.

Three-layer structure: Training Data — Simulation Evaluation — Physical AI Map
Three-layer structure: Training Data — Simulation Evaluation — Physical AI Map
Training chain diagram: Human video → Simulation → Real robot with data evolution flywheel
Training chain diagram: Human video → Simulation → Real robot with data evolution flywheel
Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

Embodied AI3D ReconstructionAmapSpatial DataPhysical AISimulation EvaluationRobot Training
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.