🇩🇪 Germany's Robot Revolution · MEC524 Case Study 2026

Control Theory
in the Real World

From Munich to Leipzig, German engineers are deploying autonomous robots that run on the same mathematical foundations you are learning this semester. This module traces four pillars of control engineering — Laplace, PID, Stability, and State Space — through the machines reshaping modern industry.

20,000+ Agile Robots deployed worldwide
90,000 Components moved by BMW humanoid
4 Core control topics — one story
Scroll to explore
Nov 2025
Agile ONE launched — humanoid enters Bavaria production
Transfer function modelling
Dec 2025
BMW AEON first deployment — Leipzig plant test run
PID wheel & joint control
Mar 2026
DFKI announces 6G robot concept — Hannover Messe
Network stability margins
Mar 2026
Agile Robots × Google DeepMind — Gemini Robotics partnership
Physical AI = state estimation
Chapter 1 · Laplace & Transfer Functions

Every joint has a transfer function

Before Agile Robots could manufacture the Agile ONE humanoid in Bavaria, their engineers had to model every arm segment, every joint, every sensor — mathematically. That process begins with Laplace.

🤖 Agile ONE · Munich, Germany
Control Concept at Work

Agile ONE's dexterous hand contains force-torque sensors in every joint. To design a controller for precise gripping, engineers must first derive the transfer function of each finger segment — mapping applied torque as input to angular position as output.

⬡ Block Diagram — Agile ONE Robot Arm Joint (Simplified)
R(s) Σ − Controller G_c(s) Plant: Motor + Arm Joint Sensor H(s) = τ(s) Y(s) Feedback: Force/Torque sensor reading G_OL(s) = G_c(s) · G_p(s) Closed-loop: Y(s)/R(s) = G_OL / (1 + G_OL · H)
Laplace Transform
ℒ{f(t)} = F(s) = ∫₀^∞ f(t)e^(−st) dt
Converts time-domain ODE of a robot joint (2nd order) into an algebraic expression in s. This is how engineers analyse Agile ONE's arm without solving differential equations numerically.
Transfer Function — DC Motor
G_p(s) = K / [s(Js + b)(Ls + R)]
The plant model for a motor-driven joint. J = inertia, b = damping, L = inductance, R = resistance. All robot arm joints map to a variation of this structure.
Closed-Loop TF
T(s) = G_c · G_p / (1 + G_c · G_p · H)
The relationship between desired position R(s) and actual position Y(s) for Agile ONE's finger. The denominator determines the stability and speed of response.
🦾
Real-World Connection — Agile ONE
Agile ONE's hand contains fingertip and force-torque sensors in every joint, enabling it to grip delicate components in electronics manufacturing. Each sensor closes a feedback loop — which is exactly the block diagram you just saw. The AI model doesn't replace the control loop; it sits on top of it, setting the reference signal R(s) based on the task objective.
Discussion Questions
Q1
Why do engineers use the Laplace domain instead of directly solving differential equations for robot arm dynamics?
In the Laplace domain, differential equations become algebraic expressions. This makes it possible to derive the transfer function, analyse the system's poles and zeros, and design a controller — all symbolically, without numerical simulation. For a complex humanoid with 71 degrees of freedom like Agile ONE, this simplification is essential for systematic controller design.
▶
Q2
Agile ONE's AI model is "trained on Europe's largest industrial dataset." How does this relate to the plant model G_p(s) in a traditional control loop?
A traditional G_p(s) is derived analytically from physics (Newton's laws, circuit equations). An AI-learned model approximates the same relationship from data. In modern Physical AI systems, both coexist — the analytical model provides safety bounds and initial tuning, while the learned model refines behaviour for complex, unmodelled tasks like pick-and-place of irregular objects.
▶
Q3
If the denominator of the closed-loop transfer function T(s) contains a root in the right-half plane, what physical behaviour would you observe in Agile ONE's gripper?
A right-half plane pole corresponds to an unstable response — the output grows unboundedly. Physically, the gripper would oscillate with increasing amplitude and potentially damage the part being handled, or cause the arm to thrash uncontrollably. This is why stability analysis (Chapter 3) is performed before deploying any controller.
▶
Chapter 2 · PID Controllers

Three terms that drive AEON

BMW's AEON robot at Plant Leipzig moves at 2.5 m/s on wheeled legs across a factory floor. Its navigation, joint tracking, and battery swap precision are all governed by controllers — and PID remains the workhorse of industrial robotics.

🏭 BMW AEON · Leipzig, Germany
Control Concept at Work

AEON uses 22 integrated sensors including time-of-flight, infrared, SLAM cameras, and microphones. Speed control of its wheel drives, torque tracking of its arm joints, and smooth stopping at workstations all rely on PID loops running in real-time — some at 1 kHz or higher.

⬡ PID Control Loop — AEON Wheel Speed
v_ref Σ − e(t) PID Controller K_p + K_i/s + K_d·s P: respond I: eliminate D: dampen Motor Drive Wheel + Body v(t) Encoder Wheel encoder → measured velocity feedback
⬡ Interactive — Tune AEON's Wheel Speed Controller
Setpoint (v_ref = 1.0 m/s)
Actual velocity v(t)
PID Time Domain
u(t) = K_p·e(t) + K_i∫e(τ)dτ + K_d·ė(t)
The control signal sent to AEON's motor driver. K_p corrects the current error, K_i eliminates steady-state offset over time, K_d dampens overshoot — critical when AEON must stop precisely at a workstation.
PID in s-Domain
G_c(s) = K_p + K_i/s + K_d·s
The Laplace form of PID. Notice how the integral term (1/s) adds a pole at the origin — this is what forces zero steady-state error in type-1 systems like AEON's velocity loop.
Ziegler–Nichols Quick Tune
K_p = 0.6K_u K_i = 2K_p/T_u K_d = K_p·T_u/8
One of the classic tuning methods. BMW's engineers would use model-based auto-tuning on AEON, but the Z-N rules provide the initial starting point before simulation refinement.
⚡
Real-World Connection — BMW AEON Leipzig
AEON can autonomously swap its own battery in 23 seconds, enabling round-the-clock operation. The docking precision required for battery connection relies on a tight PID position loop — millimetre accuracy at the docking interface. Too much K_p causes oscillation and mechanical stress on the connector. Too little means it never docks cleanly. Tuning this PID is the difference between a robot that works and one that breaks its own battery port.
Discussion Questions
Q1
In the simulator above, what happens when you increase K_p without adjusting K_d? Relate this to AEON stopping at a workstation.
High K_p with low K_d causes overshoot — the system responds quickly but oscillates around the setpoint before settling. For AEON stopping at a workstation, this would mean the robot overshoots its position target, potentially colliding with equipment. The derivative term K_d senses the rate of change and applies a braking force to damp the oscillation before it causes damage.
▶
Q2
Why is the integral term K_i important for AEON operating in high-voltage battery assembly, where gravity and friction loads change constantly?
Gravity and friction introduce constant disturbances (steady-state errors) that a pure P controller cannot eliminate. The integral term accumulates the error over time and increases the control effort until the error is zero. This ensures AEON holds precise position even when holding heavy battery modules against gravity — the I term "remembers" the bias and compensates for it automatically.
▶
Chapter 3 · Stability Analysis (Bode / Root Locus)

When the network becomes the controller

Germany's DFKI is building the next frontier: robots controlled not by onboard processors, but by their environment — via 6G networks. Shifting intelligence to the infrastructure introduces communication delays. And in control engineering, delay is the enemy of stability.

📡 DFKI 6G Robots · Hannover Messe 2026
Control Concept at Work

In DFKI's concept, the robot sends sensor data to a cloud AI over 6G, which processes it and returns a control signal. This introduces a time delay τ in the feedback path. Even a small delay of 2–5 ms can destabilize a high-bandwidth control loop. Gain and Phase margins tell us exactly how much room we have before instability.

⬡ 6G Networked Control Loop — With Communication Delay
Physical Robot e^(−τs) 6G delay τ Cloud AI Intelligence (DFKI concept) e^(−τs) return delay Actuator u(t) apply Total loop delay = 2τ (uplink + downlink) Phase lag = −2τω → reduces phase margin → may destabilise Need: GM > 6 dB PM > 30°–45°
⬡ Bode Plot Concept — Effect of 6G Delay on Phase Margin
Magnitude (dB) 0 dB +40 −40 −20 dB/dec ω_gc Phase (°) −180° −90° PM ≈ 45° PM → 0° (with delay!) Without delay With 6G delay τ
Phase Margin
PM = 180° + ∠G(jω_gc)
The additional phase lag the system can tolerate before going unstable. For DFKI's 6G robot, the communication delay adds phase lag ΔΦ = −ω·2τ. As frequency increases, this can push PM below zero — instability.
Gain Margin
GM = −20 log|G(jω_pc)| dB
How much gain the loop can absorb at the phase crossover frequency ω_pc before oscillating. Network-controlled robots require sufficient GM to account for variable latency in real deployments.
Delay Effect on Phase
e^(−τs) → ∠ = −τω rad Phase loss = −τω × (180/π) °
A 5 ms round-trip delay at 100 rad/s bandwidth causes 28.6° of phase loss. If the original PM was only 30°, the system becomes marginally stable — explaining why 6G latency specification is a control engineering problem, not just a network one.
📡
Real-World Connection — DFKI 6G at Hannover Messe 2026
DFKI's vision of robots "controlled by their environment" is elegant — but every control engineer in the room immediately asks: what is the latency? 6G promises sub-millisecond latency, which is precisely what makes this concept feasible. With τ < 1 ms round-trip, the phase loss at typical robot bandwidths (10–50 Hz) is less than 3.6° — safely within standard stability margins. DFKI's approach is a masterclass in designing the control architecture around the communication constraints.
Discussion Questions
Q1
Calculate: if the 6G round-trip delay is τ = 2 ms and the robot joint has a bandwidth of ω = 50 rad/s, how much phase margin is consumed by the delay?
Phase loss = τ × ω × (180/π) = 0.002 × 50 × 57.3 = 5.73°. For a typical PM requirement of 45°, this leaves 39.3°, which is still acceptable. However, at higher bandwidths like 200 rad/s, the same delay consumes 22.9° — this is why high-bandwidth joints (like fast finger movements) cannot be network-controlled without significant local compensation.
▶
Q2
On a Bode plot, how do you distinguish between a system that is "conditionally stable" vs "absolutely stable"? Why does this matter for DFKI's concept?
An absolutely stable system has a single phase crossover and positive PM at all gains. A conditionally stable system becomes unstable at low gains — the phase crosses −180° at multiple frequencies. For DFKI's networked robot, if the system is conditionally stable, a brief network hiccup (dropping gain suddenly) could paradoxically cause instability — a dangerous scenario on a factory floor. This is why DFKI's robots likely include local fallback controllers.
▶
Chapter 4 · State Space & Modern Control

Physical AI is just state estimation

When Agile Robots partnered with Google DeepMind, they weren't replacing control engineering — they were extending it. The Gemini Robotics foundation model does what a Kalman Filter does: it estimates an internal state from noisy, partial observations, and uses it to plan the next action.

🧠 Agile × DeepMind · Physical AI
Control Concept at Work

A Physical AI robot like Agile ONE must estimate the full state of the world — object positions, forces, its own pose — from partial sensor observations. This is the Kalman Filter / State Observer problem. The state-space framework describes systems with multiple inputs and outputs simultaneously — essential for a 71-DOF humanoid.

⬡ State Space Representation — Agile ONE Full-Body Control (Simplified)
u(t) State Equation ẋ(t) = Ax(t) + Bu(t) A: system B: input matrix ∫ dt x(t) states Output Equation y(t) = Cx(t) + Du(t) y(t) A (system) State Observer / Kalman Filter x̂(t) = Ax̂ + Bu + L(y − Cx̂) Estimates states not directly measurable
State Equation (MIMO)
ẋ(t) = A·x(t) + B·u(t) y(t) = C·x(t) + D·u(t)
A handles a system with any number of states simultaneously. For Agile ONE with 71 DOF, the state vector x contains all joint angles, velocities, and forces — far beyond what a single transfer function can describe.
State Feedback / LQR
u(t) = −K·x(t) min J = ∫(x'Qx + u'Ru)dt
Linear Quadratic Regulator minimises a cost function balancing performance (Q) and control effort (R). This is how BMW tunes AEON's whole-body balance — aggressive enough to be fast, conservative enough not to tip over.
Kalman Filter (Observer)
x̂ = Ax̂ + Bu + L(y − Cx̂) L = P·C'·R⁻¹
Estimates the hidden states from noisy sensor outputs. Agile ONE cannot directly measure the force on an object it's gripping — but its Kalman observer estimates it from motor current and joint torque — exactly what enables the "delicate touch" the company claims.
⬡ The DeepMind Connection — AI as a High-Level State Estimator
Classical State Space
  • State vector x: joint angles, velocities
  • Known A, B, C, D matrices
  • Kalman Filter estimates x̂(t)
  • LQR controller u = −Kx̂
  • Works well for known, linear dynamics
Gemini Robotics (AI Layer)
  • State: scene semantics + object poses
  • Learned from 20,000+ industrial demos
  • Foundation model estimates world state
  • Outputs reference trajectory R(s)
  • Works for unmodelled, complex tasks
Key Insight: The AI does not replace the PID or state-space controller — it runs above it, setting the reference signal. The classical control loop still runs at 1 kHz ensuring safety and stability. The AI operates at a slower planning rate (10–50 Hz), deciding what to do. Classical control decides how to do it precisely.
🧬
Real-World Connection — Agile × Google DeepMind (March 2026)
The partnership signed 4 days ago (24 March 2026) integrates DeepMind's Gemini Robotics into Agile ONE's production systems. Gemini Robotics is a foundation model — trained on data from Agile's 20,000 deployed industrial robots. When it generates a motion plan for a task, it is — in state-space terms — computing a reference trajectory x_ref(t). The onboard LQR or PID controllers then track that trajectory. You are learning the math that keeps this system from crashing.
Discussion Questions
Q1
Why is state-space representation preferable to transfer functions when designing control for Agile ONE's full body (71 DOF)?
Transfer functions work well for SISO (single-input, single-output) systems. A 71-DOF humanoid has dozens of joint torques as inputs and hundreds of sensor outputs — a MIMO system. The state-space formulation handles this naturally through matrix equations, enables modern design tools like LQR, and directly supports the Kalman Filter for state estimation. A transfer function matrix for 71 DOF would be intractable.
▶
Q2
The Kalman Filter gain L is computed from P·C'·R⁻¹. What does increasing R (measurement noise covariance) do to the observer's behaviour — and relate this to Agile ONE's force sensors in a noisy factory?
A larger R means "I trust the sensor measurements less." The Kalman gain L decreases, so the observer relies more on the model prediction Ax̂ + Bu and less on the sensor correction. In a noisy factory environment with vibrations from stamping machines, Agile ONE's force sensors will have high noise — the Kalman filter would be tuned with a larger R to smooth out the noise, at the cost of slower response to real changes in grip force. This is a fundamental engineering trade-off.
▶
Q3
If you were an engineer at Agile Robots, at what point in the control hierarchy would you insert the Gemini AI model? Draw or describe the layered architecture.
A reasonable architecture: (Top) Gemini AI operates at ~10 Hz — receives camera/LIDAR data, outputs a reference pose or trajectory x_ref(t). (Middle) LQR/State Feedback at ~100 Hz — tracks the reference using full state feedback, computes optimal joint torques. (Bottom) PID current/torque loops at ~1 kHz — directly control motor currents to achieve the commanded torques. Each layer runs at its own timescale, with slower layers setting references for faster layers — this is called a cascaded or hierarchical control architecture.
▶
Synthesis

Four topics. One machine.

Germany's robot revolution is not magic. It is the direct application of the four pillars of control engineering — applied systematically, layered carefully, and deployed with rigour. Here is how they connect:

01
Laplace & Transfer Functions
Agile ONE — Bavaria
Every finger joint, every motor is a transfer function. Before any robot can be controlled, it must be modelled in the s-domain. This is the language in which engineers write down what a machine does.
02
PID Controllers
BMW AEON — Leipzig
The workhorse of every robot ever built. AEON's wheels, arms, and battery dock all run PID loops. Understanding P, I, and D is not a stepping stone — it is a lifelong engineering tool used everywhere from satellites to surgical robots.
03
Stability Analysis
DFKI 6G Robots — Hannover
When intelligence moves to the network, delay enters the loop. Gain margin and phase margin are the quantities that tell engineers whether a networked system will remain stable — or oscillate itself apart. 6G's promise is a control engineering specification first.
04
State Space & Modern Control
Agile × DeepMind — Global
Physical AI does not replace state space — it extends it. The Kalman Filter, LQR, and state observer are the mathematical skeleton of every modern autonomous system. AI sits above this layer, not instead of it. You are learning the foundation that AI is built on.
A Note from Your Lecturer

The engineers who designed AEON's 6G stability margins, Agile ONE's Kalman observer, and BMW's PID torque loops all sat in a classroom like yours and worked through the same derivations. The mathematics you are learning this semester is not academic preparation for an industry — it is the industry. Germany's €5.4 billion robotics companies run on transfer functions and phase margins. Build the foundations well.