RAYEN: Imposition of Hard Convex Constraints on Neural Networks
RAYEN: Imposition of Hard Convex Constraints on Neural Networks
This research presents RAYEN, a framework for enforcing hard convex constraints on neural network outputs or latent variables. Unlike prior methods that rely on soft constraints or iterative corrections, RAYEN guarantees constraint satisfaction at all times with minimal computational overhead (e.g., 8 ms for 1K quadratic constraints). The authors validate the approach on learned locomotion policies for a quadruped robot in both simulation and real-world experiments, demonstrating speedups of 20 to 7468 times over state-of-the-art algorithms while maintaining near-optimal costs. Source code is available at https://github.com/leggedrobotics/rayen.
The paper introduces RAYEN, a framework that guarantees hard convex constraint satisfaction in neural networks during both training and testing with negligible computational overhead, validated on simulated and real-world quadruped locomotion tasks.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
overhead_time_ms
8
View original evidence
ear matrix inequality (LMI) constraints, with negligible overhead compared to unconstrained networks. For instance, it imposes 1K quadratic constraints on a 1K-dimensional variable with only 8 ms of overhead compared to a network that does not enforce these constraints. An LMI constraint with 300x300 dense matrices on a 10K-dimensional variable can be guaranteed with only 12 ms additional overhead. When used in neura
Open source S4overhead_time_ms
12
View original evidence
ear matrix inequality (LMI) constraints, with negligible overhead compared to unconstrained networks. For instance, it imposes 1K quadratic constraints on a 1K-dimensional variable with only 8 ms of overhead compared to a network that does not enforce these constraints. An LMI constraint with 300x300 dense matrices on a 10K-dimensional variable can be guaranteed with only 12 ms additional overhead. When used in neura
Open source S4speedup_factor_min
20–7,468
View original evidence
l networks that approximate the solution of constrained trajectory optimization problems, RAYEN runs 20 to 7468 times faster than state-of-the-art algorithms, while guaranteeing constraint satisfaction at all times and achieving a near-optimal cost. Finally, we demonstrate RAYEN's ability to enforce actuator constraints on a learned locomotion policy by validating constraint satisfaction in both simulation and real-w
Open source S5speedup_factor_max
20–7,468
View original evidence
l networks that approximate the solution of constrained trajectory optimization problems, RAYEN runs 20 to 7468 times faster than state-of-the-art algorithms, while guaranteeing constraint satisfaction at all times and achieving a near-optimal cost. Finally, we demonstrate RAYEN's ability to enforce actuator constraints on a learned locomotion policy by validating constraint satisfaction in both simulation and real-w
Open source S5constraint_count
1,000
View original evidence
ear matrix inequality (LMI) constraints, with negligible overhead compared to unconstrained networks. For instance, it imposes 1K quadratic constraints on a 1K-dimensional variable with only 8 ms of overhead compared to a network that does not enforce these constraints. An LMI constraint with 300x300 dense matrices on a 10K-dimensional variable can be guaranteed with only 12 ms additional overhead. When used in neura
Open source S4variable_dimension
1,000
View original evidence
ear matrix inequality (LMI) constraints, with negligible overhead compared to unconstrained networks. For instance, it imposes 1K quadratic constraints on a 1K-dimensional variable with only 8 ms of overhead compared to a network that does not enforce these constraints. An LMI constraint with 300x300 dense matrices on a 10K-dimensional variable can be guaranteed with only 12 ms additional overhead. When used in neura
Open source S4matrix_size_rows
300
View original evidence
ear matrix inequality (LMI) constraints, with negligible overhead compared to unconstrained networks. For instance, it imposes 1K quadratic constraints on a 1K-dimensional variable with only 8 ms of overhead compared to a network that does not enforce these constraints. An LMI constraint with 300x300 dense matrices on a 10K-dimensional variable can be guaranteed with only 12 ms additional overhead. When used in neura
Open source S4matrix_size_cols
300
View original evidence
ear matrix inequality (LMI) constraints, with negligible overhead compared to unconstrained networks. For instance, it imposes 1K quadratic constraints on a 1K-dimensional variable with only 8 ms of overhead compared to a network that does not enforce these constraints. An LMI constraint with 300x300 dense matrices on a 10K-dimensional variable can be guaranteed with only 12 ms additional overhead. When used in neura
Open source S4
- code: Available Artifact link Open source S6
- paper: not_reported
- weights: not_reported
- dataset: not_reported
- code: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ear matrix inequality (LMI) constraints, with negligible overhead compared to unconstrained networks. For instance, it imposes 1K quadratic constraints on a 1K-dimensional variable with only 8 ms of overhead compared to a network that does not enforce these constraints. An LMI constraint with 300x300 dense matrices on a 10K-dimensional variable can be guaranteed with only 12 ms additional overhead. When used in neura
Open source S4
l networks that approximate the solution of constrained trajectory optimization problems, RAYEN runs 20 to 7468 times faster than state-of-the-art algorithms, while guaranteeing constraint satisfaction at all times and achieving a near-optimal cost. Finally, we demonstrate RAYEN's ability to enforce actuator constraints on a learned locomotion policy by validating constraint satisfaction in both simulation and real-w
Open source S5
orld experiments on a quadruped robot. The code is available at https://github.com/leggedrobotics/rayen
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org