JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models
JoyNexus is pitched as shared infrastructure for post-training VLA models without giving each tenant a private GPU stack.
The paper describes a service that separates training, inference, and environment components behind APIs. Tenant-specific modules, optimizers, rollout records, and policy versions are kept isolated while global queues schedule shared resources. Its group-batching method lets compatible heterogeneous VLA samples share a backbone forward pass. The authors report simulated and embodied-scenario results showing lower aggregate GPU time and better utilization than isolated single-tenant execution. ArXiv · AI/CL/LG's note
The paper describes a service that separates training, inference, and environment components behind APIs. Tenant-specific modules, optimizers, rollout records, and policy versions are kept isolated while global queues schedule shared resources. Its group-batching method lets compatible heterogeneous VLA samples share a backbone forward pass. The authors report simulated and embodied-scenario results showing lower aggregate GPU time and better utilization than isolated single-tenant execution. ArXiv · AI/CL/LG's note
score 4