Researchers from Fudan University and Singapore Management University report that a robot arm completed 50 of 60 attempts across four precision tasks when its controller received two extra views rendered from a simulation of its workspace. Using the same controller and real cameras without those views, the arm completed 11. The results come from an Oct. 8 arXiv preprint.
SpatialHarness keeps a digital version of the robot and objects in its work area, then renders views from angles chosen to make important alignments easier to see. The arm still uses a real wrist camera and a fixed third-person camera. The added views come from a MuJoCo simulation updated from camera images and the arm’s measured movement; they do not require extra camera hardware.
The tests used a Franka Research 3 arm controlled by GPT-6 Astra. Each setup was tried 15 times on block stacking, plug insertion, placing a holed piece over a Tower of Hanoi peg, and opening a drawer before putting a block inside. SpatialHarness raised success on every task: stacking went from 7 of 15 trials to 15; plug insertion from 4 to 10; peg placement from 0 to 15; and the drawer task from 0 to 10.
The action controller stayed the same, but preparing the digital scene and updating it during robot movement add work. The authors report that synchronization takes extra time, especially when the gripper picks up or releases an object. The study covers one arm and four tasks involving rigid or articulated objects; it does not test smaller models or deformable objects.
Sources
- SpatialHarness paper and experimental results, arXiv:2610.12457v1, posted Oct. 8, 2026.
- Image: Ims, “Franka Emika3.jpg,” Wikimedia Commons, CC BY-SA 4.0. Used unmodified.
