Amap Launches Phys AI Data, a One-Stop Spatial Data Platform for Physical AI Training and Deployment

On July 8, Alibaba Group's mapping unit Amap officially launched Phys AI Data, described as the industry's first one-stop spatial data platform purpose-built for physical AI training and application. The system comprises two core products: Phys AI Foundry for simulation-based training and Phys AI Map for real-world deployment.

Physical AI's transition into open-world environments faces a critical bottleneck in data. Training a robot to operate in real-world conditions requires both extensive raw training material covering open scenarios and the ability for robots to understand their spatial surroundings. Phys AI Data addresses both pain points through an integrated approach.

Phys AI Foundry functions as a dedicated data factory for physical AI training, combining three data pathways: real-world robot collection, synthetic data generation, and simulation reconstruction. Leveraging Amap's multi-channel data acquisition capabilities across indoor and outdoor environments, Phys AI Foundry has accumulated millions of task-specific robotic action datasets, immersing robots in real-world observation-action-feedback loops from their earliest training stages.

The platform also includes an automated synthetic data engine that can batch-produce multimodal video with long temporal sequences, strong causal relationships, and high interactivity, targeting complex long-horizon tasks that real-world collection struggles to cover. For simulation-based reconstruction, Phys AI Foundry uses Amap's massive spatiotemporal data to create high-fidelity 1:1 recreations of real environments, providing safe, controllable training grounds for edge cases like extreme weather.

Phys AI Map serves as what Amap calls a "spatial memory brain" for robots. Unlike humans who can navigate using common sense, robots require machine-readable spatial semantics. Phys AI Map builds a dedicated spatial semantic library that enables models to recognize environmental features such as construction zones or poorly lit areas, elevating maps from coordinates to inferable environments.

Phys AI Map has constructed a fine-grained spatial road network and achieved 90% coverage of high-traffic indoor scenes, enabling seamless cross-floor and cross-area navigation. At key decision points like intersections and escalators, the system places multimodal visual anchors that allow robots to self-locate in milliseconds using their onboard cameras.

Phys AI Data is now fully open to the embodied intelligence industry with standardized API access and customized data services, enabling companies to leverage Amap's data and spatial capabilities without costly self-collection or map-building.