Grounding data, unified.
Robot trajectories, egocentric interactions, and general visual data are converted into supervision with answers that can be checked against annotations.
✧ PERCEPTION FOR PHYSICAL INTELLIGENCE
ACE-Brain-0.6
Scaling Spatial Grounding
for Physical Foundation Models
ACE-Brain Team01 / THE IDEA
Better grounding.
A stronger perceptual foundation.
TL;DR
Can diverse, verifiable supervision bring a 9B model to frontier-level spatial grounding? ACE-Brain-0.6 fine-tunes Qwen3.5-9B on approximately 1.5 million samples and exceeds GPT-6 Astra on four of seven public grounding benchmarks, with a higher average under the manuscript’s evaluation protocol.
75.41Public benchmark Avg-7+9.69 points over Qwen3.5-9B
93.42%RobotGrounding macro score14 scoring groups, held-out robot scenes
1.54MTraining samplesApproximately 3.41M question–answer pairs
Coverage that scales.
With the backbone and training recipe fixed, successive data releases expand both sample count and the range of grounding scenarios.
Small model, precise targets.
A 9B model improves on its base checkpoint across six of seven public benchmarks and leads the compared models on RobotGrounding.
02 / FROM RECORDINGS TO SUPERVISION
Teaching a model where.
Grounding connects a language instruction to a physical scene. Our pipeline turns diverse observations into points, boxes, counts, and spatial judgments.
One model. Five answer forms.
Coordinates are integers normalized to [0, 1000]. A shared text interface covers target identification, complete instance coverage, spatial reasoning, and interaction grounding.
[x, y][[x, y], …][x₁, y₁, x₂, y₂]countyes / no / none- Backbone
- Qwen3.5-9B
- Training
- Answer-token SFT · 1 epoch
- Vision encoder
- Frozen
- Sequence length
- 16,384 tokens
- Peak learning rate
- 1 × 10⁻⁵
03 / THE EVIDENCE
Grounding, measured.
A shared evaluation protocol across general-purpose and embodied models. Explore the public benchmarks or the held-out robot-scene evaluation.
Seven-benchmark average
Selected models · higher is better
Avg-7 is the unweighted mean of the seven public benchmark scores. RobotGrounding is excluded.
Bold: best score across all compared models. Scores use the displayed precision of the manuscript. Each benchmark uses its own official metric.
Evaluation protocol & remaining gaps
All models use the same harness and benchmark-specific prompt, temperature 0.2, and a 128-token output budget. PixMo-Count uses 1,024 tokens. Closed models are evaluated through their APIs with the same questions and scorer.
ACE-Brain-0.6 trails Astra on PixMo-Count, RefSpatial, and PointArena. The largest gap is RefSpatial: 71.84 versus 80.14. Relative to Qwen3.5-9B, PixMo-Count decreases from 90.35 to 86.96. These grounding evaluations do not establish end-to-end robot manipulation success.
04 / THE SCALING STUDY
More coverage.
Stronger spatial perception.
Across releases from 0.26M to 1.55M samples, the seven-benchmark average rises from 69.3 to 75.4.
+2.4Avg-7 points per doubling
over the evaluated range
over the evaluated range
The backbone and recipe stay fixed. Sample count and supervision coverage grow together; the main series does not isolate data volume alone.
05 / EXPLORE THE BENCHMARK
Language meets the robot’s view.
RobotGrounding contains 2,591 questions on held-out robot manipulation frames. Explore actual benchmark samples and their reference annotations.
BENCHMARK EXAMPLE
REFERENCE TARGET
Reference annotations from the benchmark, not live model predictions.
Explore 14 question types
2,591Questions
14Scoring groups
HumanBox annotations
Held-outRobot episodes
Benchmark composition & scoring
Tasks cover referring, small objects, task targets and destinations, all-instance pointing, counting, object and attribute absence, spatial judgments, and correspondence across cameras. The overall score is an unweighted mean over 14 scoring groups.
Points must fall inside the reference boxes; every-instance answers require one point per instance. Boxes are scored by intersection-over-union, while counts and categorical answers require exact matches. Box-based point evaluation can be lenient for thin or hollow objects.
06 / THE MOTIVATION
From visual references
to geometric action.
The manuscript studies GPT-6 Astra as a reference for spatial perception. These cases illustrate why selecting the right visual evidence matters for physical action.
Grounding strengths & failure cases
Astra handles diverse object, ordinal, multi-instance, spatial, and functional queries. Fine boundaries and spatially constrained placement remain difficult: a point can be near the correct part yet outside the valid region.
07 / PROJECT RESOURCES
Explore the work.
Paper
Technical report and methodology
Release link to followCode
Project repository and updates
Release link to followModel
ACE-Brain-0.6-9B model repository
Release link to followData
Grounding data and benchmark
Release link to follow08 / CITATION
Build on this work.
Manuscript citation. Publication metadata will be updated upon release.
@misc{acebrain2026grounding,
title = {ACE-Brain-0.6: Scaling Spatial Grounding
for Physical Foundation Models},
author = {{ACE-Brain Team}},
year = {2026},
note = {Manuscript}
}