Success Guided Sampling
An adaptive sampling method for large-scale parallel reinforcement learning training. By estimating the current success rate of each task configuration and reweighting the reset distribution, it concentrates training near the frontier of policy capability, thereby improving the efficiency of experience utilization.
Timeline (last 90 days)
Success Guided Sampling (SGS) is an adaptive sampler whose core idea is to concentrate RL training on configurations near the frontier of the policy's competence.
SGS does not change the underlying RL algorithm, nor does it introduce complex structural priors; instead, it adjusts the way the reset distribution is used, biasing sampling toward configurations where the policy has a chance of success but has not yet stably mastered.
The paper's experiments used up to 2^20 (more than one million) parallel environments.
At the million-scale parallel environment level, SGS enabled RL to solve challenging tasks that previous methods could not conquer, including multi-terrain quadruped locomotion and contact-rich assembly.
The paper distilled the learned manipulation policies into RGB-based policies and completed zero-shot transfer of several challenging assembly tasks on real hardware.
All Facts (5)
Success Guided Sampling (SGS) is an adaptive sampler whose core idea is to concentrate RL training on configurations near the frontier of the policy's competence.
50%UnverifiedSGS does not change the underlying RL algorithm, nor does it introduce complex structural priors; instead, it adjusts the way the reset distribution is used, biasing sampling toward configurations where the policy has a chance of success but has not yet stably mastered.
50%UnverifiedThe paper's experiments used up to 2^20 (more than one million) parallel environments.
50%UnverifiedAt the million-scale parallel environment level, SGS enabled RL to solve challenging tasks that previous methods could not conquer, including multi-terrain quadruped locomotion and contact-rich assembly.
50%UnverifiedThe paper distilled the learned manipulation policies into RGB-based policies and completed zero-shot transfer of several challenging assembly tasks on real hardware.
50%