# Control Systems and Their Limitations

In this section, we step from kinematics to control. We first show how to reason with velocities (differential inverse kinematics, diff-IK), then close the loop with feedback, and finally summarize where classical pipelines struggle in practice.

## Differential Kinematics: A Smarter Approach

Instead of solving for joint positions directly, we can work with **velocities**:

### The Key Insight
If we know the relationship between joint velocities and end-effector velocities, we can control motion more smoothly:

$$\dot{p} = J(q) \dot{q}$$

Where $J(q)$ is the **Jacobian matrix** - the relationship between joint and task space velocities.

### Differential IK Solution
Given a desired end-effector velocity $\dot{p}^*$, find joint velocities:

$$\dot{q} = J(q)^+ \dot{p}^*$$

Where $J(q)^+$ is the **pseudo-inverse** of the Jacobian.

## Adding Feedback Control

Open-loop tracking is brittle under modeling errors and disturbances. We close the loop by feeding back the tracking error.

*Dealing with moving obstacles requires feedback control.*

Real environments are **dynamic and uncertain**. We need feedback to handle:
- **Modeling errors** - Our equations aren't perfect
- **Disturbances** - Unexpected forces or obstacles  
- **Sensor noise** - Measurements have uncertainty

### Feedback Control Solution

Combine desired motion with error correction:

$$\dot{q} = J(q)^+ (\dot{p}^* + k_p \Delta p)$$

Where $\Delta p = p^* - p(q)$ is the position error.

> [!TIP]
> Start with small $k_p$ and increase gradually while monitoring oscillations. Use a watchdog (safety stop) and saturate commands to keep the system within safe limits.

## Why Classical Approaches Struggle

With differential reasoning and feedback, many tracking tasks are solvable—on paper. In practice, the system still breaks under real-world complexity for the reasons below.

*Four key limitations of dynamics-based robotics approaches.*

### 1. **Integration Challenges**
Classical pipelines are built from **separate modules**:
- Sensing → State Estimation → Planning → Control → Actuation

**Problems:**
- Errors compound through the pipeline
- Brittle when any component fails
- Hard to adapt to new tasks or robots

### 2. **Limited Scalability**  
Traditional methods struggle with:
- **High-dimensional sensor data** (cameras, LIDAR)
- **Multi-task scenarios** (each task needs custom planning)
- **Multi-modal integration** (vision + touch + proprioception)

### 3. **Modeling Limitations**
Real-world physics is complex:
- **Contact dynamics** - Hard to model precisely
- **Deformable objects** - Beyond rigid-body assumptions
- **Friction and compliance** - Difficult to characterize

### 4. **Ignoring Data Trends**
Classical methods don't leverage:
- **Growing robotics datasets** - Millions of demonstrations available
- **Cross-robot learning** - Insights from other platforms
- **Community knowledge** - Decentralized data collection

## The Learning Alternative

To address these limitations, we contrast a classical modular pipeline with an end-to-end learning policy.

**Classical Robotics Approach:**

```
Perception → State Estimation → Planning → Control → Actuation
```

**Challenges:**
- Each module needs expert tuning
- Errors compound through pipeline  
- Hard to adapt to new tasks/robots
- Requires precise world models

**Learning-Based Approach:**

```
Raw Sensors → Neural Network → Actions
```

**Benefits:**
- Learn from data - Use demonstrations and experience
- End-to-end training - Optimize the entire pipeline together  
- Generalize across tasks - Share knowledge between different objectives
- Adapt to new robots - Transfer insights across platforms

This is the promise of **robot learning**!

> [!TIP]
> **The Best of Both Worlds:** Modern robot learning often combines classical insights with learning. For example one can combine learning with safety constraints from control theory
>
> Pure learning vs pure classical is a false dichotomy - hybrid approaches have had their successes

---

## Key Takeaways

- Classical robotics relies on explicit mathematical models and expert knowledge
- Forward kinematics is straightforward, but is viable only in quite simple scenarios. Inverse kinematics is more general, but it can be challenging to develop in practice  
- Differential kinematics works with velocities rather than positions for better control
- Classical approaches struggle with integration, scalability, modeling accuracy, and data utilization
- Learning-based methods offer solutions to these fundamental limitations
- The future lies in hybrid approaches that combine classical insights with learning capabilities

> [!TIP]
> Up next, we'll show how learning-based methods (reinforcement learning and imitation learning) absorb some of this complexity by optimizing directly from data.

## References

For a full list of references, check out the [tutorial](https://huggingface.co/spaces/lerobot/robot-learning-tutorial).

- **Feedback Systems: An Introduction for Scientists and Engineers** (2008)  
  Karl Johan Åström and Richard M. Murray  
  A comprehensive introduction to feedback control systems, covering the principles that underlie closed-loop control in robotics.  
  [Book Website](http://www.cds.caltech.edu/~murray/amwiki/index.php/Main_Page)

- **Real-Time Obstacle Avoidance for Manipulators and Mobile Robots** (1986)  
  Oussama Khatib  
  A seminal paper introducing the artificial potential field method for obstacle avoidance, demonstrating how feedback can be used for reactive control in dynamic environments.  
  [DOI:10.1177/027836498600500106](https://doi.org/10.1177/027836498600500106)