Learning-Based Strategy for Composite Robot Assembly Skill Adaptation
A new AI training strategy teaches robots complex 'peg-in-hole' tasks with modular, reusable skills.
A team of researchers including Khalil Abuibaid, Aleksandr Sidorenko, Achim Wagner, and Martin Ruskowski has published a paper presenting a novel, learning-based strategy to solve a classic robotics challenge: contact-rich assembly. The work, accepted at RAAD 2026, tackles the difficulty industrial robots face with tasks like 'peg-in-hole' assembly, where tight tolerances, friction, and uncertain contact dynamics often cause failures, especially for position-controlled manipulators.
The core innovation is a reusable, encapsulated skill-based framework that employs Residual Reinforcement Learning (RRL). Instead of training a robot from scratch for every new task, the method breaks the assembly process into composite skills with explicit pre-, post-, and invariant conditions. This modular structure allows skills to be reused across task variations. RRL then enables safe and sample-efficient adaptation by only learning small 'residual' adjustments to these predefined skills during the tricky contact phases, while the overall skill sequence remains stable.
Evaluated in a MuJoCo simulation using a UR5e robot arm with a Robotiq gripper and trained with the SAC algorithm in JAX, the proposed formulation demonstrated robust execution. By combining the structure of skill-based programming with the adaptability of reinforcement learning, this approach promises to make industrial robots significantly more flexible and easier to deploy for complex, variable assembly work without extensive retraining.
- Uses Residual Reinforcement Learning (RRL) for safe, sample-efficient adaptation of contact-rich assembly skills.
- Encapsulates tasks into modular, reusable composite skills with explicit execution conditions for flexibility.
- Demonstrated robust 'peg-in-hole' assembly in simulation on a UR5e robot, trained with SAC in JAX.
Why It Matters
This could drastically reduce programming time and increase flexibility for industrial robots in manufacturing, enabling them to handle variable, real-world assembly tasks.