Critic Architecture Matters: Dual vs. Unified Critics for Humanoid Loco-Manipulation
评论家架构的重要性:双评论家与统一评论家在人形机器人移动操作中的对比
机构 * Department of Computer Engineering(计算机工程系) ; Çukurova University(丘布亚大学)
AI总结 针对人形机器人多目标强化学习,对比统一评论家与双评论家架构,实验表明双评论家策略在到达速度、吞吐量和成功率上显著优于统一评论家,且架构选择比奖励工程影响更大。
Comments Accepted at the ICRA 2026 Workshop on Reinforcement Learning for Imitation Learning (RL4IL), Vienna. 5 pages, 2 figures. v2: corrects the unified-critic run's curriculum level (10 of 40) and per-run environment counts, adds a Confounding Factors section, and softens the causal framing; measurements unchanged. https://mturan33.github.io/critic-architecture-matters/