AIAny
Icon for item

WarpSAC: Towards the Pinnacle of Scalable Off-policy RL by Rethinking Exploration and Exploitation

Adapts off-policy RL stabilizers to the available data regime: introduces WarpSAC, a regime-aware family using Sample Weight Decay plus two regime-matched variants (WarpSAC-L and WarpSAC-A) to improve sample efficiency, wall-time learning, and sim-to-real deployment.

Introduction

Scalable parallel simulation flips the usual data assumptions of off-policy RL: instead of narrow replay coverage that demands conservative stabilizers, massive parallelism supplies broad, rapidly refreshed experience that can make some stabilizers restrictive. WarpSAC reframes the problem as regime-matching—apply different stabilizers depending on whether training is data-limited (CPU-scale) or data-abundant (GPU-parallel), while always using an age-biased sampler to focus updates on policy-relevant transitions.

Key Findings
  • Regime-dependent stabilizers: parameter projection normalization and clipped double-Q help when replay coverage is narrow but can hurt value fitting when replay is abundant; relaxing them can improve learning at scale.
  • Sample Weight Decay (SWD): an age-biased linear decay sampler that concentrates updates on recent, policy-relevant transitions; beneficial across both regimes and especially when network capacity is limited.
  • Practical variants: WarpSAC-L (Norm ON + clipped double-Q) for CPU-scale, WarpSAC-A (Norm OFF + single-Q) for GPU-parallel. Using these prescriptions, WarpSAC improves normalized score–step AUC by ~4.5% (CPU-scale) and ~23.1% (GPU-parallel), raises a UnitreeG1 task success rate from 19.8% to 96.4%, gains ~19.1% mean normalized wall-time AUC on MuJoCo Playground, and achieves ~36.4% faster sim-to-real deployment (35 vs 55 minutes) in an A800 end-to-end run.
Who it's for and tradeoffs

Great fit if you run high-throughput simulation or robot learning pipelines and need to match algorithmic stabilizers to available replay regimes rather than stacking conservative fixes. It’s useful for researchers and engineers doing large-scale parallel simulation, robot manipulation, or sim-to-real workflows who can modify replay sampling and critic/normalization choices.

Look elsewhere if you only need off-the-shelf RL on low-throughput environments without access to parallel simulators, or if you require new auxiliary networks or architectural changes—WarpSAC focuses on regime-aware stabilizer selection and sampling rather than adding model complexity.

Information

  • Websitearxiv.org
  • OrganizationsTianjin University, Shanxi University, Imperial College London
  • AuthorsZihao Wu, Hongyao Tang, Yi Ma, Huizhong Song, Pengyi Li, Yifu Yuan, Fei Ni, Jinyi Liu, Wei Wei, Jianrong Wang
  • Published date2026/08/25

More Items

Proposes Recuris, a recursive Experiential-Working Memory architecture that separates Working Memory (task progress) from Experiential Memory (skills) and uses a Meta-Agent to validation-gate localized skill updates, enabling bounded recursive skill evolution for long-horizon agents.

Converts image-level rewards into explicit intermediate targets for diffusion-model denoising via an on-policy self-distillation loop. Constructs bounded positive/negative targets around anchors from reward gradients, fits those targets with finite updates, and refreshes a behavior policy by EMA—improving aligned performance across backbones while reducing GPU hours.

Treats human annotations as oracle rollouts and separates them from on-policy baselines to improve reinforcement learning for video multimodal LLMs. Key features include a decoupled advantage estimator, sign-balanced pruning, and scalable gains across model sizes and data budgets.