Tsinghua and Infinigence AI Open-Source RLark for Embodied Intelligence

RLark is built around an embodied-runtime and task-level cross-cluster networking. It allows robots and cameras to be requested, scheduled and reused like GPUs, while separating cross-cluster traffic by task and optimizing large and small packet transmission, QbitAI reported. The platform places cloud computing, edge nodes and embodied devices into one resource system and organizes execution by complete task, so access, deployment and communication configurations from one experiment can be reused in later tasks. Operators can connect and manage devices with one command, and researchers can deploy training, inference and real-robot interaction programs to different clusters with one task configuration.

The report said RLark has completed unified management of three clusters and nearly 100 cloud-edge-device nodes, covering four models of embodied devices. Device onboarding has been cut from hours to about five minutes, and tasks can start within 10 seconds after submission. RLark has also been combined with the RLinf training framework to connect cross-region collection, training and real-robot validation into a closed loop.

To test whether cloud computing and on-site devices can operate around the same task, the team combined RLark with RLinf in a cross-region real-robot experiment between a cloud GPU cluster in Guangdong and a robot site in Beijing. The experiment involved on-site robots and cameras, cloud GPUs, and training, inference and real-robot interaction roles. RLark handled device requests, cross-cluster deployment, task instance interconnection and status collection, while RLinf handled training computation and data collaboration.

During preparation, operators connected the cloud GPU cluster and the on-site device cluster to RLark. Agents synchronized node capacity, resource information and operating status. The embodied-runtime then registered adapted dual-arm robots and cameras as resources that could be requested and scheduled by tasks. In the test environment, onboarding dropped from about one hour to five minutes. Researchers could view the location, type and availability of resources in both locations from the platform and reuse them in later experiments.

After resources were ready, researchers defined training, inference and real-robot interaction roles in one task configuration, declaring required resources, instance counts and deployment locations. For example, the training role used cloud GPUs, while the real-robot interaction role requested on-site robots and cameras, all under the same embodied task. After submission, RLark sent the configuration to the relevant clusters, and agents created execution instances and established cross-cluster communication links. In the test environment, the task started on the platform within 10 seconds after submission.

The experiment ran continuously for about 36 minutes and advanced to 323 global training steps. On-site robots and cameras generated interaction data, which was sent back to the cloud for RLinf training; updated policies were then used in subsequent real-robot interaction, forming a collection-training-validation loop. Researchers could view the status of each role from one task entry and trace anomalies through execution instances, nodes, devices and logs. When changing devices, adjusting role scale or algorithm parameters, they could modify and reuse the task configuration.

RLark also conducted dedicated communication tests for task-level cross-cluster networking. The platform uses virtual addressing, a gVisor user-space network stack and SSH secure tunnels to establish communication channels for execution instances distributed across cloud and on-site locations. The platform maintains routing, tunnels and forwarding relationships. In a constrained network environment, RLark's large-packet single-stream throughput was about 49% higher than the VPN solution used in the test, and small-packet single-stream throughput was about 14% higher. In a high-bandwidth environment, large-packet single-stream throughput approached 2 Gbps. Compared with EasyTier, RLark reduced stutter caused by network transmission during cross-region real-robot tasks, according to QbitAI.

The platform uses a control-plane and data-plane separated architecture. Users manage resources and submit tasks through a web console, API or command-line tool. The control plane maintains resource information, task configurations and communication relationships, while agents in each cluster handle local task deployment, resource synchronization and status reporting. Four core technologies support the platform: an embodied device plugin and runtime, task-level cross-cluster communication optimization, declarative task orchestration, and task-level observability.

The embodied device plugin and runtime turn robots and cameras into recognizable, requestable and schedulable resources in a cloud-native system. Device plugins discover and register adapted embodied devices and provide resource information to Kubernetes; cluster agents synchronize resource information and operating status to the control plane. In a single task configuration, researchers can declare the types and counts of GPUs, robotic arms and cameras, and the platform organizes their use. The task-level cross-cluster networking connects execution instances in different clusters through virtual addresses and SSH secure tunnels. RLinf can localize traffic processing, organizing observation data processing, inference and real-robot interaction at the edge to reduce the bandwidth cost of sending all raw observation data back to the cloud. Declarative task orchestration uses Job, Task and Worker models: a Job represents the entire task, Tasks describe training, inference and real-robot interaction roles, and Workers are execution instances. Workflow can organize multiple Jobs with stage dependencies. Task-level observability links Jobs, Tasks, Workers, nodes, devices, events and logs, and provides Web Terminal and TensorBoard entries to check runtime environments, device access and training processes.

RLark has opened its user interface, API, backend services, task orchestration, cross-cluster interconnection and embodied device runtime. The open-source repository is at github.com/RLinf/RLark, with documentation at rlark.readthedocs.io and a quick-start guide at github.com/RLinf/RLark#quick-start. The report said the release aims to lower the threshold for using embodied infrastructure so more teams can deploy and use it.