Futian Laboratory

“He Zhang — Tencent Robotics X / Futian Laboratory (Corresponding Author). The corresponding author and likely the principal researcher driving this work.”

“On standard benchmarks (in-distribution), GTA-VLA performs comparably to the best existing systems... GTA-VLA (81.2%) outperforms GR00T-N1 (57.1%) on SimplerEnv-Bridge”

“GTA-VLA scores 61.4% average success vs. 52.3% for the next-best competitor (X-VLA) and 7.3% for π0.5.”

“GTA-VLA introduces a third input channel — sparse visual cues like a click (affordance point), a drawn box, or a sketched path — that directly conditions the robot's reasoning chain before action is generated.”

“This work was done during an internship at Futian Laboratory — notable that a strong system came from an intern project, suggesting the lab structure is productive”

“Qing Lian | Futian Laboratory | Equal co-first author on the core system design | Co-equal contribution on all architecture and experiment decisions”

“GTA-VLA is built directly on Qwen3-VL-2B, chosen for 'strong multimodal understanding and spatial grounding capabilities' (Section 3.1)”