Recoleta
GitHub
Source note

JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models

Authors
Haoran Sun, Wentao Zhang, Junyang Hua, Hedan Yang, Yongjian Guo, Yifei Zhang, et al.
Source
arXiv
Published
Collection
Embodied AI
Robot Foundation ModelVision Language ActionRobot Data ScalingSim2realGeneralist Robot Policy
Open arXivMarkdown source

Summary

JoyNexus presents a service-oriented platform for multi-tenant post-training of vision-language-action models. It shares resident VLA backbones across supervised fine-tuning, reinforcement learning, rollout, and evaluation workloads while isolating each tenant’s trainable modules and state.

Problem

  • VLA post-training spans heterogeneous robot embodiments, simulators, datasets, action schemas, and objectives, making infrastructure management difficult for individual users.
  • Exclusive GPU allocation and fixed card-hour accounting can waste resources during rollout, data loading, evaluation, or other bursty phases.
  • Existing service-oriented post-training systems mainly target language models and do not directly provide VLA-specific environment interaction, policy synchronization, and evaluation workflows.

Approach

  • JoyNexus separates the Training Model Service, Inference Model Service, and Environment Service behind APIs coordinated by a Master Service.
  • A shared base VLM remains resident, while tenant-specific action modules, optimizer states, checkpoints, and policy versions occupy isolated slots and remain separately addressable.
  • High-level APIs compile SFT, RL, and evaluation specifications into service workflows; lower-level APIs allow users to compose custom algorithms.
  • Global Training and Inference Queues schedule concurrent tenant workloads, while fault isolation, monitoring, and elastic rollout scaling support long-running jobs.
  • Group batching combines heterogeneous samples with a compatible model-facing prefix so multiple tenants can reuse one shared-backbone forward pass.

Results

  • Workload simulation and a group-batching pipeline in a realistic embodied scenario reportedly reduce aggregate GPU time and improve service utilization compared with isolated single-tenant execution.
  • The provided text reports no numerical values for GPU-time reduction, utilization, latency, throughput, or comparison workload size, so the magnitude of the efficiency improvement cannot be determined from this excerpt.
  • The paper evaluates system efficiency rather than proposing a new VLA policy, training objective, or robot-control benchmark.

Link