On 20 October SenseTime's newsroom described work by Daxiao Robotics, which released ACE-Brain-0.6, a 9B spatial-grounding model for embodied systems. The base is Qwen3.5-9B with the vision encoder frozen; training is supervised fine-tuning on roughly one million samples drawn from about 940,000 images and 2.7 million question-answer pairs, built from more than 30 open robot-data families plus private first-person data.
Labels are produced automatically: key frames are chosen by gripper state and motion, an object vocabulary is built per scene, and Grounding DINO and SAM 3 supply boxes and regions, with checks that a counting question is generated only when instance annotations are complete and that 'target absent' questions are verified against the frame. The company reports six wins out of eight grounding benchmarks against GPT-6 Astra: Where2Place 76.55 vs 71.10, PixMo-Count 89.41 vs 88.10, RoboSpatial 74.82 vs 69.55, RoboAfford 88.51 vs 84.94, ShareRobot 49.70 vs 46.60, and its own RobotGrounding 94.13 vs 91.10; the average over the seven public benchmarks reads 74.53 against 73.68. RobotGrounding holds 2,591 questions across 14 task types, including multi-instance counting and missing-target judgement.
A controlled scaling run grew training data from 260,000 to one million samples and moved the seven-benchmark average from 69.3 to 74.5, described as close to log-linear at about 2.4 points per doubling. Adding more samples of the data types already present gained less than one point; new scenes, relations and supervision forms supplied the rest.
Our takeThe headline is a 0.85-point average lead, thinner than the run-to-run variance most of these suites carry, and no repeated runs or error bars are given — the per-benchmark gaps, Where2Place at 5.45 points especially, are the numbers with signal. The more useful result is the ablation: another 700k samples of the same shape bought under one point, while new scenes and relations bought about five. That reframes embodied data work as coverage design rather than collection volume, and anyone can falsify it by rebuilding the recipe.