Embodied AI

Embodied AI examines intelligence through an agent that senses and acts within an environment. The body is not merely an output device: its sensors, geometry, dynamics, and physical limits shape what the agent can learn and do.

Perception through action

An embodied agent can move to obtain better information. A robot may change its camera view, touch an object, or reposition a grasp. This creates an ongoing perception–action loop rather than a one-way pipeline from input to answer.

Physical and simulated embodiment

Research can use real robots or simulated agents. Simulation enables scale and controlled experiments; hardware exposes systems to contact, wear, latency, sensor noise, and unmodeled variation. Strong work treats the relationship between the two as an engineering question rather than assuming perfect transfer.

Why it matters

Embodied AI connects machine learning with robotics, control, cognitive science, and human–robot interaction. It helps frame questions about data collection, spatial understanding, memory, active perception, and learning from consequences.

LeRobot v0.6.0 makes that loop tangible for developers: record multimodal episodes, train a policy, evaluate it in simulation, deploy it on supported hardware, capture human interventions, and use corrections as new training data. The loop is useful precisely because failure and assistance remain visible. See the practical release guide for what the available software establishes, and what still requires real-robot evidence.

Embodiment also extends beyond humanoids. Our six-case production evidence comparison covers mounted construction arms, warehouse cells, bricklaying fleets, bipeds, and a wheeled dual-arm robot under one operating framework.

See how autonomy, demonstrations, and teleoperation are disclosed across real platforms in Humanoid Robots in 2026. Then continue with robot foundation models or begin the structured learning path.