Embodied AI for General-Purpose Manipulation

An open-source embodied AI project evolving through real-world deployment — from modular single-robot planning, to data-driven dual-arm manipulation, to collective adaptive intelligence across heterogeneous robots.

AIRSEAI is an open-source embodied AI project that advances through three evolutionary stages — from modular single-robot planning, to data-driven dual-arm manipulation, and ultimately to collective adaptive intelligence across heterogeneous robots operating in complex real-world environments.

_____________________________________________________________________________________

AIRSEAI: A Different Way to Build Embodied AI

Most conversations about embodied AI jump straight to a single destination: a general-purpose humanoid robot that can do anything a person can. AIRSEAI takes a different view — that embodied intelligence is unlikely to emerge fully formed in one universal robot. Instead, it develops the way biological systems do: through repeated interaction with real tasks, real data, real hardware constraints, and real deployment.

AIRSEAI follows a deliberate three-stage roadmap. Each stage is designed to solve the limitations exposed by the one before it — building reliable modular behavior first, expanding physical capability through data-driven manipulation next, and coordinating diverse robots at system scale after that.

Hosted by LF AI & Data, AIRSEAI provides a neutral, transparent, community-driven foundation where researchers, developers, hardware companies, and robot manufacturers can build interoperable embodied AI capabilities together.

____________________________________________________________________________

The AIRSEAI Ecosystem

Three sub-projects provide the compute, data, and spatial-perception infrastructure that supports the AIRSEAI roadmap across all stages.

AIRSEAI-EmbodiCore serves as the hardware compute foundation for the mainline AIRSEAI ecosystem, delivering dedicated on-chip AI acceleration and low-latency edge processing tailored for embodied intelligence. It features a specialized NPU/TPU architecture optimized for Vision-Language-Action (VLA) models alongside a microsecond-level co-processor for hard real-time motion control, kinematics, and safety. By providing hardware-level timestamp synchronization across RGB-D cameras, LiDAR, IMU, and tactile arrays, EmbodiCore supplies the high-throughput execution power required for real-world robotic operations. 

AIRSEAI-EmbodiData acts as the data infrastructure and acquisition engine for AIRSEAI, offering an end-to-end toolchain to gather, process, and expand multimodal datasets for model training. It supports multi-modal trajectory and manipulation capture through teleoperation setups, mobile devices, VR/AR, and exoskeletons, combined with automated semantic labeling and human-to-robot motion retargeting. Furthermore, EmbodiData leverages physics-based simulation and generative AI to synthesize high-fidelity corner-case data, ensuring robust model performance across diverse Sim2Real scenarios. 

AIRSEAI-EmbodiMap provides the spatial perception and semantic mapping framework for AIRSEAI agents, translating raw multi-sensor streams into interactive 3D environment models. Utilizing 3D Gaussian Splatting, NeRF, and SLAM technologies, it constructs incremental topological maps embedded with open-vocabulary language representation, enabling agents to execute long-horizon natural language commands in physical space. Its real-time dynamic filtering and incremental update engines continuously adjust to environmental changes while enabling seamless multi-robot collaborative mapping.

______________________________________________________________________________

2026 Roadmap & Milestones

________________________________________________________________________________

Get Involved

AIRSEAI welcomes researchers, developers, hardware companies, robot manufacturers, and end users who want to help make embodied AI more deployable, interoperable, and accessible.

• Explore the code: [GitHub →]
• Join the community: [TSC Mailing List  →]
• Follow the project: [Project Website →]

AIRSEAI was donated by the Shenzhen Institute of Artificial Intelligence and Robotics for Society (AIRS) to LF AI & Data as an Incubation-stage project in July 2026.