💻 computer science

Simulation-based Learning of Electrical Cabinet Assembly Using Robot Skills

This paper presents a simulation-driven approach that combines deep reinforcement learning with parameterizable robot skills and realistic physics-based joining models to successfully automate the force-controlled assembly of electrical terminals on DIN-rails, achieving high success rates in both simulation and real-world settings without additional tuning.

Original authors: Arik Laemmle, Balázs András Bálint, Philipp Tenbrock, Frank Naegele, David Traunecker, József Váncza, Marco F. Huber

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arik Laemmle, Balázs András Bálint, Philipp Tenbrock, Frank Naegele, David Traunecker, József Váncza, Marco F. Huber

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to snap a plastic electrical clip onto a metal rail, like clicking a Lego brick into place. Sounds simple, right? But in the real world, these clips are tricky. They are made of flexible plastic that bends, snaps, and sometimes fights back. If the robot pushes too hard, it breaks the clip. If it pushes too soft, it doesn't snap in. If the rail is slightly crooked or the clip is a millimeter off, a traditional robot programmed with rigid instructions would just fail and give up.

This paper is about teaching a robot to learn how to do this job by itself, using a "virtual playground" (simulation) before ever touching the real world.

Here is the story of how they did it, broken down into simple concepts:

1. The Problem: The "One-Size-Fits-All" Trap

Traditionally, programming a robot for this task is like writing a strict script for a play. You tell the robot: "Move 5cm left, push down with 10 Newtons of force." But in a factory, every electrical box is slightly different. The rails might be crooked, the clips might be from a different batch, or the temperature might have changed the plastic.
If you have to rewrite the script for every single tiny variation, it takes too long. The robot becomes too expensive to use for small batches of products.

2. The Solution: A "Video Game" Training Camp

Instead of writing a script, the researchers taught the robot using Deep Reinforcement Learning (DRL). Think of this as training a dog, but the dog is a robot and the treats are digital points.

The Virtual World: They built a physics-based video game (using a tool called MuJoCo) that looks and feels exactly like the real factory.
The Agent: The robot is the "player." It doesn't know how to snap the clip yet. It just knows it wants to get points.
The Trial and Error: The robot tries to snap the clip.
- Fail? It gets a negative score (a "boo").
- Success? It gets a positive score (a "high five").
- Over millions of tries in the computer, the robot learns the perfect way to wiggle, push, and twist to get the clip to snap in, even if the rail is crooked.

3. The Big Hurdle: The "Bendy" Problem

Here is the tricky part: Most video game engines (physics simulators) are great at simulating hard things like metal blocks or walls. They are terrible at simulating bendy things like plastic snap-hooks.
If you try to simulate a bending plastic hook in a standard game engine, the engine usually just says, "That's a solid block," and the robot learns the wrong thing.

The Fix: The researchers had to invent two new "magic spells" (mathematical models) to teach the computer how plastic bends:

The Analytical Spell: They used old-school physics math (beam theory) to calculate exactly how much force is needed to bend the plastic.
The Rigid-Body Spell: They broke the plastic hook down into tiny, invisible springs and hinges inside the computer. This made the computer think the plastic was actually bending, even though it was just a bunch of rigid parts connected by springs.

They tested these spells against real-world measurements and found they were accurate enough to train the robot.

4. The "Skill" Backpack

Instead of teaching the robot to move every single motor individually (which is like teaching a human to move every single muscle fiber), they gave the robot a backpack of pre-made skills.

Skill A: "Approach the rail."
Skill B: "Slide along the rail."
Skill C: "Pivot and snap."

The robot's job wasn't to figure out how to move its arm; its job was to figure out which skill to use and how hard to push based on what it felt. It's like teaching a chef not how to chop a carrot, but when to chop, how fast to chop, and what knife to use based on the carrot's shape.

5. The Grand Finale: From Video Game to Reality

Once the robot mastered the task in the video game, they transferred the "brain" (the learned policy) directly to a real robot arm (a UR10e) in a real factory.

The Result?

No Re-training: They didn't have to tweak the robot's settings once it arrived in the real world.
Superhuman Accuracy: The robot succeeded 100% of the time, even when the rail was crooked or the clip was in a weird spot.
Generalization: They trained it on one type of clip, and it could successfully snap in different types of clips it had never seen before.

The Takeaway

This paper proves that we don't need to be robot programmers anymore; we just need to be robot coaches. By creating a realistic virtual world where robots can practice millions of times, and by giving them the right "math magic" to understand bending plastic, we can teach robots to handle messy, real-world tasks that used to be too difficult to automate.

It's the difference between teaching a robot to follow a map, and teaching it to navigate the terrain.

1. Problem Statement

The assembly of electrical terminals onto DIN-rails in control cabinet manufacturing is a force-controlled task traditionally hindered by:

High Programming Effort: Manual programming for robots is time-consuming and often exceeds the time required for manual assembly.
Product Variability: High diversity in product variants and small batch sizes require frequent reprogramming and adaptation.
Simulation Limitations: Standard rigid-body physics engines (like MuJoCo) struggle to accurately simulate the deformation forces of snap-hooks (flexible polymer components), which are critical for force-controlled assembly. Without accurate force simulation, Deep Reinforcement Learning (DRL) agents cannot be effectively trained to handle real-world tolerances.
Skill Parameterization: While "robot skills" (predefined modules) reduce complexity, their parameters still require manual expert tuning, limiting flexibility for small batches.

2. Methodology

The authors propose a framework combining Deep Reinforcement Learning (DRL) with parameterizable robot skills within a physics-based simulation environment.

A. System Architecture

Robot Skills Framework: The system uses the pitasc framework, which abstracts robot control into hierarchical skills. The assembly task is decomposed into a sequence of skills:
1. $S_{lin}$ : Linear approach to a pre-position.
2. $S_{approach}$ : Linear movement toward the rail.
3. $S_{slide}$ : Sliding along the rail with constant contact force.
4. $S_{pivot}$ : Rotating the terminal to engage the snap-hook.
5. $S_{pivot}$ (optional): Final orientation adjustment.
Learning Agent: The agent uses DRL to select the appropriate skill and optimize its parameters (e.g., speed, force, rotation angle) based on sensor observations (relative position, orientation, and force/torque data).
Algorithms: The study employs Soft Actor-Critic (SAC) and Twin Delayed Deep Deterministic Policy Gradient (TD3) for off-policy learning.

B. Simulation & Joining Models (Core Innovation)

To bridge the gap between rigid-body simulation and deformable snap-hooks, the authors developed two types of joining models integrated into MuJoCo:

Analytical Joining Models:
- Based on beam theory (cantilever beam equations).
- Calculates lateral force ( $F_Q$ ) and joining force ( $F_J$ ) based on deflection, material properties (secant modulus), and geometry.
- Provides a computationally efficient way to approximate forces without complex deformation dynamics.
Rigid-Body Joining Models:
- Decomposes the deformable snap-hook into a series of linked mass-spring-damper sub-models.
- Uses constraints within MuJoCo to simulate the bending and deflection dynamics.
- Includes variations: 1-DoF translational ("Slide"), 1-DoF rotational ("One-hinge"), and 2-DoF rotational ("Two-hinge").

C. Training Strategy

Domain Randomization: To ensure robustness, the simulation randomizes the DIN-rail's position ( $\pm 5$ mm) and orientation ( $\pm 3^\circ$ ), as well as sensor noise.
Reward Function: Based on the distance to the target position and orientation, penalizing deviations.
Sim-to-Real Transfer: Policies trained in simulation are transferred directly to a physical UR10e robot system without additional fine-tuning.

3. Key Contributions

Novel Joining Models: The development and evaluation of both analytical and rigid-body models to accurately simulate snap-hook deformation forces in a standard physics engine, enabling realistic force-controlled training.
Skill-Based DRL Pipeline: A successful integration of task-level DRL with modular robot skills, allowing the agent to learn not just how to move, but which skill parameters to adjust for specific uncertainties.
Zero-Shot Sim-to-Real Transfer: Demonstrating that policies trained entirely in simulation can achieve high success rates on physical hardware with no additional tuning, even for untrained terminal variants.
Comprehensive Evaluation: Extensive testing on a physical testbed covering various terminal types, positions, and significant rotational/translational deviations.

4. Results

Simulation Performance:
- The trained agents achieved up to 100% success rates in simulation.
- The agents successfully generalized to new terminal types and positions not seen during training.
- The Analytical Model showed a mean absolute error (MAE) of <3 N after optimization.
- The Rigid-Body Models (specifically the 2-DoF rotational) provided the most accurate force profiles but were computationally slower (Real/Sim ratio < 1), whereas the analytical and overlap-based methods were faster than real-time.
Real-World Performance (Physical Robot):
- Transferred policies achieved 100% success rates in core and transition areas on the physical UR10e robot.
- Robustness: The system successfully compensated for DIN-rail positional deviations up to $\pm 5$ mm and orientation deviations up to $\pm 6^\circ$ (industrial typical inaccuracies are usually much smaller).
- Generalization: Policies trained on one terminal type (e.g., WAGO 2002-1201) successfully assembled different, stiffer terminal types (e.g., WAGO 2102-1301).
- Failure Modes: Failures were primarily due to extreme rotations ( $>8^\circ$ ) or physical slippage of the terminal in the gripper, rather than algorithmic failure.

5. Significance and Impact

Scalability for Small Batches: This approach significantly reduces the engineering time required to automate new product variants. Instead of manual programming, only the skill parameters (geometry, material) need to be defined, and the agent learns the optimal execution strategy.
Bridging the Sim-to-Real Gap: By accurately modeling the non-linear forces of snap-hooks, the paper demonstrates that simulation-based learning is viable for complex, force-sensitive assembly tasks previously considered too difficult for automated learning.
Modularity: The combination of DRL with the pitasc skill framework offers a scalable architecture where new skills can be composed, and learning can be applied to higher-level decision-making rather than low-level joint control.
Future Applications: The methodology is applicable to any assembly task involving deformable parts (snap-fits, cables, soft robotics), paving the way for fully automated, flexible manufacturing cells in Industry 4.0.

In conclusion, the paper validates that combining physics-accurate simulation of deformable parts with skill-based reinforcement learning creates a robust, adaptable automation solution capable of handling the high variability and precision requirements of modern electrical cabinet manufacturing.