GSBR: Geometric Support Behavior Regularization for Online RL
GSBR: Geometric Support Behavior Regularization for Online RL
This research introduces Geometric Support Behavior Regularization (GSBR), a technique designed to enhance online reinforcement learning by penalizing actions based on their distance from a pretrained base policy's support. Unlike previous methods that were either too restrictive or limited improvements, GSBR allows the policy to focus on high-value modes within the support while permitting improvements outside it when the value gained outweighs the penalty. Evaluated across 19 tasks on four robotic benchmarks, GSBR significantly improves performance, raising the average success rate from 33% for the base policy to 95%, and achieving the highest average success rate and throughput among all compared methods.
The paper introduces Geometric Support Behavior Regularization (GSBR), a method that penalizes actions based on their distance from a pretrained base policy's support to balance stability and improvement in online reinforcement learning, achieving an average success rate of 95% across 19 tasks on four robotic benchmarks compared to 33% for the base policy.
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
Reported success rate
33%
View original evidence
quires only sample access to the base policy. Across 19 tasks on four robotic benchmarks, GSBR raises the average success rate from 33% for the base policy to 95%, achieving the highest average success rate and throughput among all compared methods, with about 23% higher normalized throughput than the next-best RL method. Project website and open-source code at https://charlesxu0124.github.io/gsbr/ .
Open source S5Reported success rate
95%
View original evidence
quires only sample access to the base policy. Across 19 tasks on four robotic benchmarks, GSBR raises the average success rate from 33% for the base policy to 95%, achieving the highest average success rate and throughput among all compared methods, with about 23% higher normalized throughput than the next-best RL method. Project website and open-source code at https://charlesxu0124.github.io/gsbr/ .
Open source S5tasks
19
Reported trials
View original evidence
quires only sample access to the base policy. Across 19 tasks on four robotic benchmarks, GSBR raises the average success rate from 33% for the base policy to 95%, achieving the highest average success rate and throughput among all compared methods, with about 23% higher normalized throughput than the next-best RL method. Project website and open-source code at https://charlesxu0124.github.io/gsbr/ .
Open source S5robotic benchmarks
4
View original evidence
quires only sample access to the base policy. Across 19 tasks on four robotic benchmarks, GSBR raises the average success rate from 33% for the base policy to 95%, achieving the highest average success rate and throughput among all compared methods, with about 23% higher normalized throughput than the next-best RL method. Project website and open-source code at https://charlesxu0124.github.io/gsbr/ .
Open source S5
- code: Available Artifact link Open source S5
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
quires only sample access to the base policy. Across 19 tasks on four robotic benchmarks, GSBR raises the average success rate from 33% for the base policy to 95%, achieving the highest average success rate and throughput among all compared methods, with about 23% higher normalized throughput than the next-best RL method. Project website and open-source code at https://charlesxu0124.github.io/gsbr/ .
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org