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

An embodied AI experiment site is often assembled from multiple roles: on-site devices such as robots and cameras, cloud compute, training and inference programs, and the communication environment connecting them. The project description relayed by Machine Heart likens this collaboration to an airport control tower — aircraft fly on their own, but takeoff and landing require unified coordination with runways, flight paths and ground resources. As robots move from single units to fleets, the cost of connecting devices one by one, deploying tasks separately, repeatedly configuring networks and troubleshooting failures individually gets amplified.

RLark's entry point is not the single matter of "connecting devices in," but organizing scattered compute, devices and execution programs to run under the same task. According to official statements, the platform centers on a self-developed embodied device runtime and task-level cross-cluster networking: the former uniformly manages the runtime environments of devices such as robots and cameras, while the latter performs cross-cluster communication isolation and optimization for different tasks and traffic characteristics.

It should be noted that the above positioning and architecture descriptions all come from an article the project side published via Machine Heart, and are source_claim; this article has not obtained independent third-party verification, so expressions such as "unified management" and "scheduling like GPUs" should be understood as project goals and self-reported capabilities, not externally confirmed conclusions.

The officially disclosed scale is: RLark has completed unified management of 3 clusters and nearly 100 cloud-edge-end nodes, covering 4 models of embodied devices. In the test environment, embodied device onboarding was cut from a traditional 1 hour to about 5 minutes; when resources are already connected and ready to run, a task can start on the platform side within 10 seconds of submission. These numbers are the most concrete quantitative basis in this article, but they all come from the project side's own account and are explicitly limited to the "test environment."

The cross-region test combined RLark with the reinforcement learning infrastructure framework RLinf, running between a cloud GPU cluster in Guangdong and a robot site in Beijing. On-site devices handled interaction and data collection, cloud GPUs handled training compute, and the trained policies were then used for subsequent real-machine interaction. Officially, the experiment ran continuously for about 36 minutes, advancing training to 323 global training steps, and fully ran through the entire chain of device request, cross-cluster scheduling, real-machine data return, cloud training and policy update.

The dedicated communication test gave another set of numbers: under constrained network conditions, RLark's large-packet single-stream throughput improved by about 49% and small-packet single-stream throughput by about 14% over the VPN approach used in the test; under high-bandwidth conditions, large-packet single-stream throughput approached 2 Gbps. The article also claims that compared with the EasyTier approach, RLark significantly reduced stuttering caused by network transmission during cross-region real-machine task runs. These comparison numbers are likewise source_claim, and the test conditions, baseline configuration and reproducibility were not elaborated in the article.