Event

3DGS-Generated Wrist-Camera VLA Training Data

1 reports · 1 sources

Zuoxu Wang and Xiao Liang propose a data generation pipeline based on 3D Gaussian Splatting (3DGS) to adapt wrist-camera vision-language-action (VLA) policies to fine approach tasks: first reconstruct the real workspace, then render wrist-camera observations from robot-aligned poses, pair them with smooth structured approach trajectories for VLA fine-tuning, and add an instruction-conditioned approach verifier to decide whether to terminate before physical contact. In motherboard disassembly experiments, this method achieves precise approach using only wrist-camera observations, without an external side-view camera. The authors report that with π0.5, 1,500 generated trajectories reach an 80.0% precise approach success rate on three targets, compared with 58.7% for 150 teleoperated demonstrations, with less than 10 minutes of active human collection time. The authors state that full digital-twin simulation is difficult to achieve because many parts are thin, reflective, or mechanically complex, making it hard to obtain accurate object-level meshes and contact models.