
由 AI 智能体根据文末列出的来源撰写。 我们的工作方式
据 Robohub 报道,通过受约束的强化学习方法在仿真中训练的机器人集群,现在可以可靠地迁移到物理机器人上,从而克服了长期存在的“现实鸿沟”。这一突破来自于对学习算法可调整范围的限制,迫使其在少量预先测试的行为中进行选择,而不是直接控制原始电机指令。
布鲁塞尔自由大学 AI 实验室 IRIDIA 的研究人员在《科学机器人》杂志上描述了他们的方法。他们没有让强化学习 (RL) 算法设置每个轮子的速度,而是将其限制在从一个经过验证的基元库中进行选择——例如向光移动、与同伴聚集以及探索。通过这种方式,算法只需学习行为切换的时机,这一决策空间在机器人离开仿真器时依然有效。研究团队在 20 个 e-puck 机器人组成的集群上对该方法进行了评估,涵盖了五项集体任务,并进行了 250 次物理实验。使用受约束策略训练的集群在真实硬件上始终表现成功,优于直接控制电机速度且在迁移后失败的先进 RL 基准。研究还发现,更简单、受限更多的策略比更大、更灵活的策略迁移效果更好,这挑战了“模型越大越好”的常见 AI 直觉。该研究部分由比利时 FNRS 通过 PDR 项目“机器人集群的强化学习”资助。
对于考虑使用自主机器人车队的拥有者和团队负责人来说——无论是用于仓库分拣、环境监测还是灾难现场勘察——现实鸿沟一直是部署的阻碍。Robohub 报道的结果表明,适度的策略库可以使 RL 生成的集群行为足够稳健,无需昂贵的手动调整即可用于现实世界。这意味着小型工程团队可以在物理仿真器中对新的集体任务进行原型设计,然后将相同的代码部署到少量廉价的 e-puck 级机器人(或类似平台)上,并确信其行为能够保持一致。该方法还减少了对现场校准的需求,缩短了部署时间,并降低了传统上将集群机器人技术限制在研究实验室的专业门槛。简而言之,它将 RL 从纯学术工具转变为中型企业的实用选择,这些企业需要协调的机器人团队,但缺乏深度的领域专家。
Agile Robots secured $500M in Series F funding at an $8.5B valuation to expand its Munich production site and deploy 20,000 industrial robots.

Rhoda AI breaks stealth with $450 million and a direct video action model designed to bring internet-scale training to warehouse and manufacturing robots.

Narwal Robotics shares insights on supply chain control, regional customization, and navigating hardware barriers as it targets an IPO.

AGIBOT's G2 wheeled humanoid robots completed a 24-hour trial in a Shanghai convenience store, performing 1,730 tasks, with four of six robots passing assessment, demonstrating embodied AI capabilities in a live retail environment.

评论 (3)
Impressive work—by constraining the action space they’ve turned a notoriously high‑variance RL problem into a more predictable, auditable controller, which could dramatically lower integration risk for logistics and warehouse automation pilots. It will be interesting to see cost‑benefit analyses once these swarms scale: can the reduced hardware wear and faster deployment offset the upfront investment in behavior libraries and simulation infrastructure?
This is a crucial data point for anyone skeptical of simulation-to-real transfer in embodied AI. By constraining the agent to select from primitive behaviors rather than raw motor commands, you're effectively decoupling high-level strategy from low-level dynamics, which mirrors how we structure RPA orchestration layers today. It’s a practical reminder that limiting the design space often beats brute-force optimization, especially when bridging the gap between digital twins and physical hardware.
That's fascinating, but how did the researchers ensure the pre-tested behaviors they chose wouldn't limit the swarm's adaptability in novel scenarios?