Tsinghua and Genesis-Cloud Open-Source RLark, a Cloud-Native Platform Orchestrating Embodied AI
机器之心 · wechat · 2026-09-24
Tsinghua University and Embodied/Genesis team (无问芯穹) have open-sourced RLark, a cloud-native orchestration platform—described as an "air traffic tower" for embodied AI—that unifies the management of robots, cameras, cloud GPUs, and training/inference programs.
Key capabilities:
- embodied-runtime: abstracts physical robots and cameras into schedulable, GPU-like cloud-native resources.
- Task-level cross-cluster networking: virtual addressing, gVisor userspace networking, and SSH tunnels boost large-packet single-flow throughput 49% over a tested VPN setup (near 2Gbps in high-bandwidth environments).
- Declarative orchestration: a Job/Task/Worker model lets one config define resources and placement for training, inference, and real-robot interaction roles.
- Task-level observability for tracing failures from job down to device.
In a cross-region real-robot experiment linking a Guangdong GPU cluster with robots in Beijing, device onboarding dropped from hours to 5 minutes, tasks launched within 10 seconds, and a 36-minute run completed 323 global training steps—closing the data-collection/training/real-robot-validation loop. The platform now manages 3 clusters, 100 nodes, and 4 robot models. The related RLinf-USER paper appears at RSS 2026 (arXiv:2602.07837). Repo: github.com/RLinf/RLark.
More from Embodied
- TJ-FlyingFish: Chinese team builds 1.6kg amphibious drone that flies and dives for 40 minutes — lukas_m_ziegler · 2026-09-24
- Skydio unveils F10 Lightrunner: 100 mph drone with 30-mile radius, auto-launched by MegaDock — Scobleizer · 2026-09-24
- Daily AI brief: Meta glasses VR, GPT-6 voice plugins, Gemini TTS, Opus 5.5 — testingcatalog · 2026-09-24
- Berkeley's Do as I Do turns everyday human videos into dexterous robot hand training data — micoolcho · 2026-09-24
- Meta Mocked for 'Inventing the Tamagotchi' as Users Say They Want One — jonstephens85 · 2026-09-24
- Robots fold laundry and write code, but camera calibration still takes hours — RexDouglass · 2026-09-24