Embodied AI MaaS shifts robot models to cloud-based post-training
新智元 reports that 原力灵机 upgraded its embodied-intelligence MaaS service to support SFT, DAgger and RECAP post-training in one platform. It says the service lets developers upload demonstration data, train models, deploy inference and adapt models to different robot bodies without building their own training infrastructure. The report also relays IDC figures for China’s MaaS market and vendor claims about deployment success rates, but it does not provide the underlying reports or independent validation.
Cloud-based embodied model services could lower the technical barrier for robot makers, while combining multiple post-training methods addresses the task-specific accuracy required for robots to move from demonstrations into deployed workflows.
ASI启示录 2026-11-05 12:07 北京
新智元报道
2026年,大模型卖Token这门生意,已经卷到让人看不懂了。
谷歌在今年I/O大会上甩出一个数字:每个月处理3200万亿个Token,一年翻了7倍。
国内更猛。IDC上周刚发的报告显示,今年上半年中国MaaS市场规模冲到106亿元,同比增长近9倍;Token消耗量2857.9万亿,同比增长超过430%。
#原力灵机具身智能MaaS#Physical Intelligence Pi API#Dream API#DM0.5#NVIDIA Thor#Google#IDC