Daxiao Robotics' ACE-Brain-0.6, described on SenseTime's newsroom on 20 October 2026, is a 9B spatial-grounding model built for embodied systems. The reported headline is a 0.85-point lead over GPT-6 Astra on a seven-benchmark average, 74.53 against 73.68. That figure is the least informative part of the announcement. The consequential material sits underneath it: an ablation in which roughly 740,000 additional samples of the same shape bought less than one point, while new scenes and relations supplied the rest. One of those results is a leaderboard claim that sits inside plausible run-to-run noise. The other is a design statement about where grounding capability actually comes from.
The method is compact enough to restate exactly. ACE-Brain-0.6 fine-tunes Qwen3.5-9B under supervised fine-tuning with the vision encoder frozen, on roughly one million samples drawn from about 940,000 images and 2.7 million question–answer pairs. The corpus combines 30-plus open robot-data families with private first-person data. Labels are generated rather than collected by hand: key frames are selected by gripper state and motion, each scene gets an object vocabulary, Grounding DINO and SAM 3 supply boxes and regions, and guards restrict counting questions to frames with complete instance annotations while verifying "target absent" questions against the frame itself.
Freezing the vision encoder is the most consequential engineering choice in that recipe, and it is a cost decision as much as a modelling one. A frozen tower removes the largest activation and optimizer-state burden from training, keeps the dominant share of the parameter count out of the gradient path, and lets the team iterate on data and supervision without re-tuning a perceptual front end. What it constrains is equally clear. Grounding precision is capped by whatever the encoder's features already encode, and any shift toward first-person robot imagery has to be absorbed by the trained side of the network, which can re-weight representations but cannot rebuild them. The announcement does not state the encoder's size, training compute, or whether an adapter is trained between tower and language model, so the magnitude of that constraint cannot be estimated from the summary alone.
The lead is thinner than it looks
An 0.85-point average gap is not a robust separation on suites of this kind. Benchmark scores in this range commonly move by more than a point across prompt formats, sample subsets and scoring conventions, and the announcement reports no repeated runs, no seeds and no error bars. Taken as a single draw, the average should be read as a tie with a directional hint. The per-benchmark gaps carry more signal precisely because they are larger than plausible formatting drift: Where2Place at 76.55 against 71.10 is a 5.45-point separation, RoboSpatial 74.82 against 69.55 is 5.27, RoboAfford 88.51 against 84.94 is 3.57, ShareRobot 49.70 against 46.60 is 3.10, RobotGrounding 94.13 against 91.10 is 3.03, and PixMo-Count 89.41 against 88.10 is 1.31. The pattern is a model that is clearly better at placement and spatial relations and roughly level on counting.
The average also hides a hole. Eight benchmarks are reported with six wins, but the average is taken over seven, and the excluded or unnamed entries are not identified. If the six named scores are among the seven averaged, arithmetic puts the unstated seventh near 48.6 for the stated mean to hold, against roughly 64.4 for the comparison system — a deficit near 16 points on a suite the summary does not name. That is an inference from the published numbers, not a reported figure, but it matters for interpretation: a 0.85-point lead assembled alongside a double-digit loss is a different object from a uniform small edge. Note also that ShareRobot at 49.70 is low in absolute terms for both systems, which suggests a suite where neither is close to saturation and where a 3.10-point gap is worth less than it appears.
What the scaling run licenses
The controlled run is the strongest evidence in the release, and it is a substitution experiment rather than a scale demonstration. Growing the training set from 260,000 to 1,000,000 samples moved the seven-benchmark average from 69.3 to 74.5, described as close to log-linear at about 2.4 points per doubling; two doublings at that rate predict 4.8 points against the observed 5.2, which is internally consistent. The decomposition is what makes it useful. Additional samples of data types already present gained less than one point. New scenes, relations and supervision forms supplied the remainder. The axis that mattered was coverage, not volume.
That reframes embodied data work as a design problem about what the corpus contains rather than how much of it there is, and it makes the claim falsifiable in a way leaderboard entries usually are not. The pipeline is described concretely enough — frame selection by gripper state and motion, per-scene vocabularies, detector-plus-segmenter labelling, explicit guards — that another group could rebuild it and check whether the same substitution produces the same slope. The obvious limit is that 2.4 points per doubling is a fit along one axis over roughly two doublings. Whether the relation holds past one million samples, or on suites other than the seven averaged, is not stated.
The label pipeline is part of the model
Every automated stage is a filter, and its failure modes propagate into the benchmark numbers. Per-scene vocabulary coverage determines which objects can be asked about at all, so classes outside the vocabulary become invisible rather than merely hard. Detector misses become candidate "target absent" negatives, and the verification guard exists precisely because such errors would teach the model that present objects are missing; the guard's recall therefore bounds label quality on the category most likely to be silently corrupted. Segmenter region quality sets the ceiling on box precision. Key-frame selection driven by gripper state biases the corpus toward manipulation moments and away from navigation and passive observation. The guards are the right instinct — they convert silent noise into skipped samples — but skip rates, per-stage yield and any human audit are not stated, so the residual noise floor is unknown.
RobotGrounding deserves separate scrutiny because it is self-built: 2,591 questions across 14 task types, including multi-instance counting and missing-target judgement, which are exactly the two categories the guards govern. A 94.13 score there is partly a statement about consistency with the team's own labelling conventions. The summary gives no inter-annotator agreement, no cross-check against human labels, and no statement of whether evaluation scenes are held out from training, so comparability with the public suites is unresolved. A practitioner should ask for the per-benchmark table including the unnamed seventh suite, repeated runs with variance, explicit held-out-scene separation, and latency and memory at deployment — a 9B model plus vision tower on robot compute is a real constraint, and no inference cost, throughput or memory figure is given. The durable contribution here is not the 0.85 points. It is the substitution result and the pipeline description, which together specify a claim someone can try to break.
Sources: https://www.sensetime.com/cn/news/ace-brain-0-6-scaling-20261020-1419