Recoleta
GitHub
Source note

LabVLA: Grounding Vision-Language-Action Models in Scientific Laboratories

Authors
Baochang Ren, Xinjie Liu, Xi Chen, Yanshuo Liu, Chenxi Li, Daqi Gao, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot Foundation ModelScientific Lab AutomationSimulation Data GenerationCross Embodiment
Open arXivMarkdown source

Summary

LabVLA targets robot execution of written laboratory protocols, where current vision-language-action policies fail on lab instruments, liquids, and multi-step bench workflows. It pairs a synthetic lab data engine with a two-stage policy so a pretrained VLM can map lab instructions to robot actions across embodiments.

Problem

  • Existing VLA policies are mostly trained on household and tabletop data, so they do not learn lab-specific objects, contact precision, or protocol workflows.
  • Real laboratory data is costly to collect because it needs specialized instruments, safety checks, and domain supervision.
  • The same lab protocol must run on different robot embodiments with different cameras, grippers, workspaces, and action spaces.

Approach

  • RoboGenesis builds lab scenes in simulation from generated 3D assets, validated layouts, and configurable robot profiles.
  • It composes long-horizon workflows from atomic skills such as pick, pour, press, stir, open, and close.
  • It applies domain randomization after workflow validation so the same protocol can vary in lighting, clutter, camera pose, object appearance, and placement without changing task semantics.
  • It exports only successful rollouts as LabEmbodied-Data with multicamera observations, robot state, actions, and step-level annotations.
  • LabVLA trains a Qwen3-VL-4B-Instruct backbone first with FAST action tokens, then with flow matching and a DiT action expert under stop-gradient knowledge insulation.

Results

  • RoboGenesis generates a LabAssetLibrary with 2,947 annotated assets and 1,000+ textures, then uses them to build 10,000 laboratory scenes.
  • The supported robot pool covers 16 robot platforms, including single-arm, bimanual, and mobile-manipulator setups.
  • The paper reports that composite workflows with more than 20 skill steps still achieve collection success rates above 75%.
  • On LabUtopia, LabVLA claims the highest average success rate among all evaluated baselines in both in-distribution and out-of-distribution settings.
  • The excerpt does not give the exact LabUtopia success numbers or baseline values, so the strongest stated result is the rank-order claim plus the reported data scale.

Link