Embodied Intelligence · Overview¶

Understand the system first

Learn why each component exists and how components work together. Use videos, diagrams and lightweight simulation before detailed derivations and training.

Robotics and embodied intelligence¶

Robotics studies sensing, motion and physical interaction. Embodied intelligence studies agents completing tasks through a body interacting with an environment. Robots are an important platform; learned policies are one family of methods.

Component	Question	Example
Body, kinematics and dynamics	What can move, how do joints relate to the end effector, and how do forces affect motion?	Arm workspace
Perception	What is present and where is the target?	Detect objects on a table
State estimation	What state explains incomplete, noisy observations?	Fuse encoders and IMU
Planning	Which states or actions should be followed?	Obstacle-avoiding path
Control	How should execution be corrected using feedback?	Adjust velocity toward a target
Policy	Which action follows from observation, state or goal?	Joint targets or end-effector displacement
Software and simulation	How are components connected and experiments repeated?	ROS 2 messages, frames and replay
Evaluation and deployment	When does the task succeed or fail?	Success, collision, timeout and latency

flowchart LR
    A[Task or language instruction] --> B[Goals and skills]
    C[Sensor observations] --> D[Perception and state estimation]
    D --> E[Planner or learned policy]
    B --> E
    E --> F[Controller and actuators]
    F --> G[Body and environment]
    G --> C

This diagram explains responsibilities. Real systems can combine components and mix learned and classical methods.

How learning methods relate¶

Method	Source of behavior	First-pass lesson
Classical planning and control	Models, objectives, constraints and control laws	A feedback baseline
Imitation / behavior cloning	Expert observation-action demonstrations	Coverage gaps can cause closed-loop failures
Reinforcement learning	Interaction and rewards	Reward design, exploration and training cost matter
VLA	Pretrained vision-language capabilities and robot action data	Understand modalities, action representation and adaptation
World model	Environment observations, states and actions	Prediction is distinct from a policy and task success

Primary course and later practice¶

Role	Selected resource	Fit	Scope
Robot motion backbone	Modern Robotics, textbook, videos and code	Builds on your linear algebra, geometry and systems experience with consistent notation	Preview and chapter introductions first; selected chapters 2–6 next; planning/control as needed
Later robot-learning practice	LeRobot tutorial	Concepts, data, training and execution share an ecosystem	Concepts first; E3 after PyTorch and action-interface checks; VLA training later

Selection evidence: Modern Robotics is required by UIUC ECE470 and appears in robotics community learning discussions. Its mathematical depth fits an engineering background; its emphasis is motion/control, with vision and learning studied later.

LeRobot is a newer official practice ecosystem with evolving tools. It is selected for workflow continuity, not presented as a decades-established course. Record versions and allow environment-debugging time.

Use official ROS 2 tutorials only for E1 software practice, and one PythonRobotics example for E2; neither is an additional full course to complete.

Exploration E0: dissect a task¶

Choose navigation to a point or an arm picking an object. Watch a demonstration and identify:

Body, sensors and available actions.
Differences between goals, observations and estimated state.
Which component chooses actions and which executes them.
Success, failure, timeout and recovery definitions.

Completion: draw a feedback loop and distinguish localization, detection, planning, control and policy learning. Full SLAM or VLA training is not required.

Later directions¶

Perception and estimation, planning and control, robot software, robot learning and edge deployment are possible specializations. Explore components on small tasks first, then select one body and task for deeper work.

Next: roadmap · Projects · Foundation overview