Tsinghua and Infinigence Open-Source RLark, a Cloud-Native Platform for Embodied AI

Tsinghua University and the AI infrastructure company Infinigence have jointly released RLark, an open-source cloud-native platform for embodied intelligence. Chinese tech outlet QbitAI framed the release as open-sourcing a new control tower for embodied AI, positioning the project around deployment rather than model research.

The headline numbers describe deployment speed. According to the source, RLark can complete robot onboarding in five minutes and start cross-cluster tasks in ten seconds. Taken together, the two figures describe the elapsed time between plugging in a new machine and running a batch of work across distributed compute.

Onboarding is an underrated part of embodied AI engineering. Getting one robot into production normally involves system images, drivers, communication links and runtime adaptation, each of which can eat hours or days. Compressing that flow into minutes means new hardware can join a usable resource pool far sooner.

Cross-cluster task launch is the other bottleneck. Training, simulation and inference for embodied systems are often spread across different compute pools, and moving work between them adds scheduling, networking and dependency-sync delays. A ten-second launch claim points to making those switches close to immediate.

The choice of a cloud-native design reflects a broader pattern in embodied AI infrastructure: using containers and cluster schedulers to absorb heterogeneous robots. The more varied the hardware, sensors and compute sources involved, the more a unified orchestration layer is worth.

The collaboration model is also notable. A university supplies research and algorithmic depth while a company brings engineering and compute experience, a combination that is becoming common in China's embodied AI sector, with open source letting outside teams validate the work quickly.

The caveat is that platform claims are settled in real deployments. Five minutes and ten seconds are operator-stated efficiency figures, and actual results depend on robot models, network conditions and cluster size. What to watch is whether the community wires RLark into production lines and lab workflows, and whether a layer of adapters and tooling grows around it.