Encord公司正在与德国神经科学初创企业Zander Labs合作，探索利用脑波传感技术来解决物理AI领域的数据瓶颈问题[1]。该项目通过在工人执行Jenga等任务期间测量其脑波活动，为机器人模型生成更优质的训练数据集[1]。脑波数据能够用于推断错误、意图和惊讶等心理状态，为模型训练提供更深层的信息维度[1]。
Encord在加州圣莱安德罗的设施内采用多模态数据收集方式，包括主从机械臂和肌肉传感器等工具[1]。该公司计划通过对比评估脑波标注数据是否能够改善机器人模型的性能表现，以此决定是否进一步扩大该技术的应用规模[1]。
物理AI训练数据的质量与成本存在显著差异。根据Encord的分析，密集标注的高质量训练数据相比低质数据的价值高出100倍，但生产成本也相应增加20倍[1]。公司高管Velmurugan指出，实现突破性进展需要的数据量相当于YouTube视频库规模的5倍[1]。
Encord is partnering with German neuroscience startup Zander Labs to explore brain wave sensing as a potential solution to the data scarcity challenge facing physical AI development [1]. The collaboration involves measuring neural activity while workers perform tasks such as Jenga, with the goal of generating higher-quality training datasets for robotic models [1]. The brain wave data is intended to capture psychological states including error detection, intent, and surprise [1].
The initiative addresses a critical bottleneck in robotic AI advancement. According to Encord, achieving a breakthrough in the field requires training datasets equivalent to five times the volume of YouTube's entire video library [1]. Meanwhile, densely annotated physical training data proves approximately 100 times more valuable than lower-quality alternatives, though producing such datasets costs roughly 20 times more [1]. At its San Leandro, California facility, Encord already employs multiple data collection modalities including master-slave robotic arms and muscle sensors to build training datasets [1]. The company plans to evaluate whether brain wave-annotated data can enhance robotic model performance before deciding whether to scale the initiative further [1].