AI

SpaceVLA Puts User Grasp and Place Anchors Into OpenVLA

Shar Hendrix 4 min read

Language is fuzzy about where to grab. “Pick up the mug and put it on the table” leaves a dozen legal grasps and landings.

SpaceVLA (arXiv:2608.05730, Aug 6, 2026) adds Visual Intent Anchors: the user points to a preferred grasp region and placement region in a Unity XR-style interface. Those points become green and blue image-space overlays on the RGB frame that an OpenVLA-7B policy (fine-tuned with LoRA) already knows how to read. Authors are from Skoltech and MWS R&D (Zinniatullina, Kolomiets, Konenkov, Altamirano Cabrera, Tsetserukou).

SpaceVLA pipeline: user anchors, anchor-conditioned OpenVLA, closed-loop Unity execution
Overview: user-authored anchors, marked RGB + language into OpenVLA, closed-loop Unity control. Source: arXiv:2608.05730.

Setup

  • 200 Unity pick-and-place demos (mug, pot, thermos, saucepan); ~150 RGB frames each
  • Split: 120 train / 80 held-out episodes; 18,000 anchor-augmented training observations
  • Policy predicts tokenized 7-DoF incremental actions (translation, orientation deltas, gripper width)
  • After grasp, the green mask drops; only blue placement stays during transport
  • Instruction for anchor policy: “Pick up the object by the green region and place it on the blue region.”
Offline dataset construction with green grasp and blue place masks on Unity demos
Offline construction of the anchor-augmented dataset. Source: arXiv:2608.05730.

Numbers (closed-loop Unity, 80 held-out episodes)

ConditionFull-task successMean grasp / place distance
Correct anchors91.25% (73/80)0.5 cm / 0.7 cm
Random anchors on objects77.5% (62/80)2.6 cm / 3.4 cm
No-anchor baseline50.0% (40/80)n/a

Correct anchors cut average grasp-and-placement action error from 3.0 cm to 0.6 cm vs random object anchors. A separate test put anchors at arbitrary workspace points (including off objects); the robot redirected toward the marks, not just the objects.

Current limitation the authors call out: anchors are fixed image-space masks under a static camera — they do not stick to moving objects or reproject under viewpoint change. Future work points to 3D world/object-relative anchors for VR/AR on real robots.

A Human’s Take

This is preference injection without full teleop, and the ablation is honest: wrong anchors hurt, no anchors collapse. Until the masks live in 3D, it is a sim-and-static-camera trick. I still want it — “green means grab here” is something a floor supervisor can understand without learning a teach pendant.

Sources