ScanArk raises pre-seed, closing three rounds in four months
ScanArk (景立构), an embodied-AI infrastructure startup spun out of Tongji University's RAIL lab, announced a pre-seed round, bringing it to three rounds — seed, seed+ and pre-seed — within four months and several tens of millions of RMB in total, with investors including Xiaomiao Langcheng, Guohai Innovation Capital, a leading humanoid robotics company and an industrial investment institution. The company says its in-house models turn real environments into "native 3DGS" scenes that carry object attributes, part relationships and interaction states, letting robots act, observe outcomes and retry inside a digital replica before validating results on real hardware. Its stated approach is to first train a base model on human demonstrations, then let robots accumulate success and failure experience in resettable, parallel interactive environments.
If simulated environments can credibly report collision, contact and dynamics outcomes, interactive learning could become an important source of robot experience beyond demonstration data — though the claims still need larger cross-scene and real-robot validation.
⾯向Physical AI的Reality-to-Sim具身智能基础设施公司景立构(ScanArk),宣布完成Pre-天使轮融资,至此,公司在4个月内已完成种子轮、种子+轮及Pre-天使轮三轮融资,累计融资数千万元人民币,投资方包括小苗朗程、国海创新资本、头部具身机器人公司及知名产投机构等。
具身智能与世界模型的发展,正面临数据供给在规模与质量上的双重挑战。互联网图像和人类第一视角视频记录了丰富的视觉信息,却往往缺少准确的空间结构、物理状态和动作反馈;真机采集能够提供真实交互经验,但成本高、效率有限,难以充分覆盖复杂任务和长尾环境。如何持续获得既贴近真实世界、又能支撑机器人学习的数据与经验,成为行业需要突破的关键瓶颈。
与此同时,GPT-6 Astra等通用大模型在复杂推理、工具使用和多步骤任务执行上的进展,正在拓展AI参与实际工作的能力边界,这也为具身智能和世界模型提出了新的命题:随着模型能力提升,训练体系需要进一步支持机器人通过行动获得反馈,在探索、试错和策略更新中持续积累经验。交互学习,有望成为具身智能下一阶段的重要学习范式。