A systems view, not a single model
There is no accepted reference architecture for embodied AGI. A useful research decomposition separates capabilities that operate at different time scales and have different reliability requirements. A multimodal foundation model may interpret goals and context; a predictive model may estimate physical outcomes; a planner may select subgoals; and a real-time controller must execute feasible actions while reacting to contact and disturbances. These components can be learned jointly, composed modularly, or combined in hybrid systems.
Six interacting layers
- Perception and grounding: fuse vision, depth, touch, audio and proprioception into object-, scene- and body-centered state estimates.
- World modeling: predict object motion, contact outcomes, agent behavior and uncertainty under candidate actions.
- Memory: maintain episodic records, semantic knowledge, maps and reusable skills with provenance and expiration rules.
- Reasoning and planning: translate goals into constraints, subgoals and long-horizon plans; replan when observations invalidate assumptions.
- Action and control: convert plans into trajectories and low-level commands while respecting dynamics, actuator limits and contact constraints.
- Assurance and oversight: monitor uncertainty, enforce safety envelopes, detect distribution shift and provide human intervention paths.
The interfaces matter as much as the modules
A system fails when its interfaces hide incompatible assumptions. A planner may request an action that the robot cannot physically execute; a world model may predict at a resolution too coarse for grasp control; a skill library may assume a different gripper or coordinate frame. Interfaces should therefore declare units, frames, timing, action semantics, confidence, preconditions and postconditions. Typed skill contracts and explicit embodiment adapters can make these assumptions inspectable.
Cross-embodiment design
A general-purpose architecture should separate shared task knowledge from body-specific execution. Shared representations can encode object relations, task phases and desired effects. An embodiment adapter maps these into the target robot's sensors, kinematics and action space. Evaluation should distinguish zero-shot transfer, calibration, few-shot adaptation and hardware-specific engineering; otherwise a result may overstate how much intelligence is actually reusable.
Evaluation and failure analysis
Measure task breadth, unseen-object performance, long-horizon completion, recovery after perturbations, calibration, latency, energy use and safety violations. Report hardware, sensors, control frequency, training data, simulator assumptions and human assistance. Include negative results and failure categories. A benchmark score on a narrow task suite is evidence of that measured capability—not proof of AGI.
A staged research roadmap
Near-term work can improve robust perception, predictive control, data-efficient adaptation and reliable skill composition. Mid-term work can test shared representations across diverse morphologies and persistent learning under changing tasks. Claims about general intelligence require broader, independent evaluation across environments and embodiments. None of these milestones guarantees a path to AGI or superintelligence.
Research sources and further reading
- Firoozi et al., Foundation models in robotics: Applications, challenges, and the future (2025)
- Ma et al., A Survey on Vision–Language–Action Models for Embodied AI (2026)
- Long et al., Learning Embodied Intelligence from Physical Simulators and World Models (2025)
- Domae et al., The Embodiment Gap in Robot Foundation Models (2026)
Explore the connected field
Embodied AGI · Robot Foundation Models · World Models · Cross-Embodiment Robotics · Superintelligence and Robotics