Zoellent AI Lab 发布 GR-SAC:用小型控制器微调冻结的 VLA 机器人策略
Zoellent AI Lab Releases GR-SAC to Refine Frozen Robot Policies With a Small Controller
如果你在玩 VLA 机器人策略,这个挺有意思:不改冻结的大模型,只加一个小控制器做残差修正,800 次测试里多跑成 5 个回合。做机器人部署、想省微调成本的开发者可以看看具体实现。
Zoellent AI Lab 推出 GR-SAC(Gated Residual Soft Actor-Critic),一种在冻结 VLA 模型之上附加小型控制器的机器人策略改进方法。在 RoboMME 基准的 800 次测试评估中,该方法相比冻结基线多完成 5 次成功回合。
其技术思路是不改动 VLA 主体权重,仅训练一个轻量的残差控制器,通过门控机制决定何时介入修正动作输出,从而避免全量微调带来的开销和灾难性遗忘。
对机器人开发者而言,这类方案可让已有 VLA 策略在特定任务上快速适配,同时保留基础模型的泛化能力,类似思路也与社区中对冻结大模型做轻量适配的研究方向一致。
Zoellent AI Lab Releases GR-SAC to Refine Frozen Robot Policies With a Small Controller
Zoellent’s GR-SAC adds a small controller to a frozen VLA, reporting five more successful episodes in an 800-test RoboMME reference evaluation.
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include GR-SAC, Zoellent, VLA, robots.
Primary source verification: published directly by pandaily with first-party authority.