ACE-Brain-0.5: A Unified Embodied Foundational Model for Physical Agentic AI

ACE-Brain-0.5: A Unified Embodied Foundational Model for Physical Agentic AI

ACE-Brain Team

ACE Robotics

Overview

ACE-Brain-0.5 is a unified embodied foundation model for Physical Agentic AI. It integrates spatial perception, decision making, embodied interaction, self monitoring, and self improvement into a single closed-loop model for physical agents.

ACE-Brain-0.5 organizes robot intelligence into five tightly coupled cognitive functions: Spatial Perception, Decision Making, Embodied Interaction, Self Monitoring, and Self Improvement.

Demo

ACE-Brain-0.5-VLA performs washing clothes tasks in a hotel laundry room.

Key Features

Unified Embodied Foundation Model

A single closed-loop model spanning spatial perception, decision making, embodied interaction, self monitoring, and self improvement.

SSR+ Training Paradigm

Extends Scaffold-Specialize-Reconcile with a Reactivate stage to unify spatial reasoning, grounding, navigation, manipulation, and progress estimation without cross-task interference.

Architecture

ACE-Brain-0.5 uses a shared embodied backbone to encode heterogeneous inputs and maintain a unified scene-and-task representation, while dedicated interfaces decode this shared state into grounding, executable planning, robot actions, and progress-estimation signals.

Evaluation Results

ACE-Brain-0.5 is evaluated across spatial QA, grounding, driving, navigation, manipulation, and progress estimation.

MindCube 86.3 Spatial mental modeling

Multi3DRef 72.4 3D language grounding

RxR SR 63.8 VLN-CE Val-Unseen

LIBERO Avg 98.2 Manipulation success rate

SimplerEnv Avg 82.3 Bridge manipulation

RBM-OOD VOC 0.96 Progress estimation

Values are from the technical report. NE and ShareRobot-Traj are lower-is-better; all other listed metrics are higher-is-better.

Spatial QA

Benchmark	GPT-5.4	Gemini-2.5-Pro	Cosmos3-Nano	RynnBrain-8B	ACE-Brain-0	ACE-Brain-0.5
VSI	52.6	43.4	54.9	71.0	63.1	62.2
MMSI	31.3	38.0	36.2	39.6	32.2	35.5
MindCube	45.3	57.6	34.8	56.6	82.1	86.3
ScanQA	78.3	67.0	60.9	60.2	97.3	99.2
SQA3D	45.8	37.3	44.2	43.3	54.5	62.6
Scan2Cap	14.0	16.8	9.2	2.8	75.2	83.3
ScanRefer	61.7	65.9	5.4	5.4	61.4	70.2
Multi3DRef	45.1	54.4	8.1	8.1	55.9	72.4
SparBench	46.1	46.2	54.8	49.8	44.2	39.7
MMSIVideo	32.8	34.5	27.2	27.5	26.4	30.4
EmbSpatial	73.2	78.7	82.9	80.0	77.8	75.9
ERQA	50.5	55.7	46.0	46.8	41.5	46.3
SAT	73.3	78.7	80.7	78.0	92.0	82.7

Grounding

Benchmark	GPT-5.4	Gemini-2.5-Pro	Cosmos3-Nano	Embodied-R1.5	RoboBrain-2.5	ACE-Brain-0	ACE-Brain-0.5
RoboSpatial	53.5	53.7	61.0	69.7	73.0	55.6	60.1
RefSpatial	15.7	36.5	53.1	54.2	60.5	26.0	55.6
PointArena	37.9	62.8	69.4	71.4	59.3	44.7	68.5
RoboAfford	29.4	15.0	84.0	80.0	74.9	56.5	75.1
ShareRobot-Traj	0.32	0.34	0.30	0.31	0.24	0.46	0.32

Driving

Benchmark	GPT-5.4	Gemini-2.5-Pro	Cosmos3-Nano	RynnBrain-8B	MiMo-Embodied-7B	ACE-Brain-0	ACE-Brain-0.5
MME-RealWorld	32.4	67.0	57.0	68.7	60.3	71.2	66.6
MAPLM	57.9	26.1	26.7	68.2	74.5	77.8	71.3
DriveAction	82.2	73.5	79.9	76.3	81.0	81.3	78.2
NuScenesQA	37.1	16.1	16.6	29.5	56.7	58.8	44.7
NuPlanQA	87.5	64.2	82.5	81.8	73.7	91.7	86.7
LingoQA	76.8	64.1	71.4	57.8	69.9	65.8	55.6

Navigation

R2R Val-Unseen

Model	NE	OS	SR	SPL
NaVid	5.7	49.2	41.9	36.5
Uni-NaVid	5.6	53.3	47.0	42.7
NaVILA	5.2	62.5	54.0	49.0
StreamVLN	5.0	64.2	56.9	51.9
NavFoM	5.0	64.9	56.2	51.2
RynnBrain-Nav	4.9	71.6	58.6	49.6
Qwen-VLA-Instruct	5.1	69.0	57.5	51.2
ACE-Brain-0.5	4.8	63.7	57.4	51.7
ACE-Brain-0.5-Specialist	4.2	67.4	62.2	56.2

RxR Val-Unseen

Model	NE	SR	SPL	nDTW
NaVid	5.7	45.7	38.2	-
Uni-NaVid	6.2	48.7	40.9	-
NaVILA	6.8	49.3	44.0	58.8
StreamVLN	6.2	52.9	46.0	61.9
NavFoM	5.5	57.4	49.4	60.2
RynnBrain-Nav	6.2	56.1	49.6	59.6
Qwen-VLA-Instruct	5.8	59.6	47.8	57.1
ACE-Brain-0.5	4.3	63.8	47.9	64.6
ACE-Brain-0.5-Specialist	4.5	63.6	51.1	67.1

Manipulation

LIBERO

Suite	OpenVLA-OFT	GR00T N1.6	pi0.5	Qwen-VLA-Instruct	ACE-Brain-0.5
Spatial	97.6	97.7	98.8	-	100.0
Object	98.4	98.5	98.2	-	100.0
Goal	97.9	97.5	98.0	-	96.0
Long	94.5	94.4	92.4	-	97.0
Average	97.1	97.0	96.9	97.9	98.2

SimplerEnv-Bridge

Object	OpenVLA	pi0	GR00T N1	X-VLA	GTA-VLA	Qwen-VLA-Instruct	ACE-Brain-0.5-VLA
Spoon	4.2	50.0	64.5	95.8	95.8	-	75.0
Carrot	0.0	41.7	65.5	75.0	87.5	-	79.2
Cube	8.3	29.2	5.5	62.5	66.7	-	75.0
Eggplant	45.8	70.8	93.0	70.8	75.0	-	100.0
Average	14.6	47.9	57.1	76.0	81.2	73.7	82.3

Progress Estimation

RBM-EVAL Progress Estimation

Split	VLAC-8B	RoboDopamine-8B	RoboReward-8B	Robometer-4B	Qwen3-VL-8B	ACE-Brain-0.5
ID Standard	0.16	0.79	0.82	0.92	0.73	0.94
ID Refined	0.19	0.55	0.51	0.78	0.28	0.80
OOD Standard	0.17	0.80	0.88	0.94	0.87	0.96
OOD Refined	0.33	0.65	0.60	0.81	0.30	0.88

BibTeX

@misc{brainteam2026acebrain05unifiedembodiedfoundational,
      title={ACE-Brain-0.5: A Unified Embodied Foundational Model for Physical Agentic AI},
      author={Brain Team and Ziyang Gong and Haoming Gu and Zehang Luo and Tianyi Zhang and Tao Tao and Yixiao Chi and Zhe Liu and Lingsi Zhu and Jingyuan Liu and Anke Tang and Songze Li and Yilun Kong and Ningjing Liu and Tianyu Zhu and Yunpeng Qing and Shuang Luo and Xiang Liu and Shi Fu and Dawei Nie and Sixiang Liu and Zhexi Wen and Feng Pan and Xiaofeng Wang and Zhi Hou and Chunxiao Liu and Xue Yang and Junchi Yan and Hengshuang Zhao and Dacheng Tao and Xiaogang Wang},
      year={2026},
      eprint={2607.04426},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2607.04426},
}