{"id":"bb715c80-3a69-4fdd-8f0b-b4ad445ed9bb","slug":"sgs","name":"Success-Guided Sampling (SGS)","description":"Success-Guided Sampling is a 2026 CoRL research method from the University of Washington and NVIDIA (Zhang, Guaman Castro, Yin, and co-authors) that changes only which task configurations PPO resets to, weighting resets toward moderate-success states so mega-scale sim RL unlocks contact-rich assembly and hard locomotion without demonstrations.","brainType":"research-model","isOpen":false,"maturityStage":"research","architecture":"Standard PPO plus an outer success-history sampler over a fixed set of task configurations (initial state, goal, environment). Resets are weighted by recent success so neither trivial nor impossible configs dominate; same reward across NIST-board assembly tasks.","reviewStatus":"reviewed","sources":[{"url":"https://sgs-rl.github.io","title":"Success-Guided Sampling project page","sourceName":"SGS project (UW / NVIDIA)","publishedAt":"2026-10-09"},{"url":"https://x.com/octi_zhang/status/2108583315544682914","title":"SGS release thread","sourceName":"X/@octi_zhang","publishedAt":"2026-10-09"}],"keyFacts":[{"label":"Contact-rich sim2real zero-shot (research-stated)","value":"CoRL 2026 SGS (UW + NVIDIA) trains PPO with success-weighted resets; UR5e meshes gears, threads nuts, inserts pegs from RGB with no demos, zero-shot from simulation."},{"label":"Scales to 1M parallel envs (research-stated)","value":"Authors report SGS keeps improving as parallel environments grow to one million, where uniform sampling and other baselines stall on locomotion and assembly."}],"aliases":["SGS","Success-Guided Sampling","sgs-rl","A Balanced Data Diet"],"collisionRisk":"low","reviewNote":null,"builtOnBrainId":null,"createdAt":"2026-10-10T20:54:26.125Z","updatedAt":"2026-10-10T21:05:50.751Z","jsonLd":{"@context":"https://schema.org","@type":"SoftwareApplication","@id":"https://registry.deploy.report/brains/sgs","url":"https://registry.deploy.report/brains/sgs","name":"Success-Guided Sampling (SGS)","alternateName":["SGS","Success-Guided Sampling","sgs-rl","A Balanced Data Diet"],"description":"Success-Guided Sampling is a 2026 CoRL research method from the University of Washington and NVIDIA (Zhang, Guaman Castro, Yin, and co-authors) that changes only which task configurations PPO resets to, weighting resets toward moderate-success states so mega-scale sim RL unlocks contact-rich assembly and hard locomotion without demonstrations.","identifier":"bb715c80-3a69-4fdd-8f0b-b4ad445ed9bb","applicationCategory":"research-model","publisher":{"@id":"https://deploy.report/#organization"}},"framework_metadata":{"framework_schema_version":"0.1.0","verification_status":"verified","maturity_stage":"research","lifecycle_state":null,"architectural_position":{"cohort":null,"sub_cohorts":[]},"within_cohort_verified_vs_claimed_pair":null,"cap_flags":[],"verification_depth":{"sources_count":2,"primary_source_types":[]}}}