Body

Degrees of freedom (DoF), end-effectors (grippers / dexterous hands), URDF descriptions, whole-body control (WBC). Morphology constrains which skills are feasible.

Perception

Active vision, vSLAM, 3D scene understanding, multimodal sensing (RGB-D, force, tactile). Unlike static image recognition, agents must move viewpoints and interact to gather information.

Decision

High-level task planning (LLM/VLM instruction decomposition) and low-level policies (IL/RL/VLA). Dual-system designs are common: a slow semantic reasoner and a fast action controller.

Action

Joint-space or end-effector control, action chunking, contact-rich dynamics. Evaluation emphasizes success rate, generalization, and safety—not language metrics alone.