ACE Robotics has released ACE-Brain-0.6, a roughly 9-billion-parameter embodied foundation model focused on visual grounding — mapping language instructions to precise locations, object parts and interaction regions in a physical scene. The company says the model is based on Qwen3.5-9B, with the vision encoder frozen and the language model fine-tuned on about one million verifiable spatial supervision samples covering roughly 940,000 images and 2.7 million question-answer pairs.
According to results released by ACE Robotics, which have not yet been independently verified, ACE-Brain-0.6 beat OpenAI's GPT-6 Astra on six of eight grounding evaluations: 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 the team's own RobotGrounding set (94.13 vs. 91.10). Across seven public benchmarks, its average score was 74.53, compared with Astra's 73.68. RobotGrounding contains 2,591 questions across 14 task types, including locating small objects, multi-instance counting, judging that a requested target is missing and matching targets across camera views.
The company also reported comparisons with Google Gemini 4 Argon, Anthropic Claude Opus 5.5 and other open embodied models, framing them as reference points for spatial perception rather than as a geopolitical contest. In controlled scaling runs that kept the architecture and training recipe fixed, expanding the training set from 260,000 to about one million samples raised the seven-benchmark average from 69.3 to 74.5, a near log-linear curve. Doubling data lifted the score by about 2.4 points on average, while simply adding more samples of the same type improved performance by less than one point. Gains came mainly from new scenes, spatial relations and supervision formats.
The data pipeline merges more than 30 open robot datasets with private egocentric footage, then uses tools such as Grounding DINO and SAM 3 to turn gripper-state and motion cues into checkable spatial labels. ACE Robotics says answers are only generated when visual evidence is complete, for example requiring full instance labels before counting questions. Public GitHub repositories under DAXIAORobotics still list ACE-Brain-0 and ACE-Brain-0.5; as of publication, ACE-Brain-0.6 weights were not listed as open.
ACE Robotics, chaired by SenseTime co-founder Xiaogang Wang, previously framed ACE-Brain-0.5 as a unified perceive-plan-act-evaluate loop. Version 0.6 narrows the bet to grounding as the link between high-level reasoning and reliable physical execution.
SEE ALSO: ACE Robotics and NTU Open-Source Puffin-World Multimodal World Model · Zhejiang University's EMem-Bench Tests Whether Embodied Agents Remember What They Saw and Tried