OmniBrain-0

OmniBrain-0

A Unified Agentic Model
for Embodied Intelligence

Xiaomi Embodied Intelligence TeamTsinghua UniversityUniversity of Macau

Technical reportCode
1unified brain
4B / 9Bmodel sizes
5task benchmarks
RSIrecursive self-improvement

Overview

OmniBrain-0 brings manipulation and navigation into one learned decision model. It interprets instructions, coordinates specialist skills, tracks task progress, and revises its plan from execution feedback.

OmniBrain-0 overview: training, agentic execution, performance, and recursive self-improvement

Available in 4B and 9B sizes, the shared brain builds on specialist controllers while learning when and how to use their skills. The RSI data flywheel turns verified successes and recovery trajectories into training experience for the next round.

Results across five benchmarks

One shared brain is evaluated on LIBERO-Pro, LIBERO-Plus, RoboCasa365, RoboTwin 2.0, and R2R-CE.

SpecialistFew-shot agentOmniBrain-0 · 9BOmniBrain-0 · RSI

LIBERO-Pro

X-VLA
3.8
π0.5
43.8
GLM-5.3-Flash
64.9
GPT-5.6-Sol
78.3
OmniBrain-0 · 9B
65.5
OmniBrain-0 · RSI
69.3

LIBERO-Plus

ABot-M0
80.5
π0.5
83.4
GLM-5.3-Flash
69.3
GPT-5.6-Sol
86.2
OmniBrain-0 · 9B
83.8
OmniBrain-0 · RSI
84.5

RoboCasa365

π0.5
16.9
RLDX-1
36.0
GLM-5.3-Flash
40.4
GPT-5.6-Sol
52.4
OmniBrain-0 · 9B
39.6
OmniBrain-0 · RSI
43.2

RoboTwin 2.0

LingBot-VLA
84.8
ABot-M0
85.57
GLM-5.3-Flash
87.0
GPT-5.6-Sol
91.2
OmniBrain-0 · 9B
88.2
OmniBrain-0 · RSI
89.4

R2R-CE

StreamVLN
56.9
DualVLN
64.3
GLM-5.3-Flash
58.0
GPT-5.6-Sol
71.0
OmniBrain-0 · 9B
65.0
OmniBrain-0 · RSI
66.0

Task success rate (%). Manipulation benchmarks report Overall. R2R-CE reports SR on full Val-Unseen for specialist policies and rand100 for agents. RSI denotes the model updated through the data flywheel.

4B, 9B, and RSI

Benchmark4B9BRSI
LIBERO-Pro62.465.569.3
LIBERO-Plus82.383.884.5
RoboCasa36533.639.643.2
RoboTwin 2.087.288.289.4
R2R-CE55.065.066.0

Training the shared brain

Three stages connect embodied knowledge, domain expertise, and cross-domain learning.

Three-stage training: embodied pretraining, domain specialization and multi-expert consolidation

Embodied pretraining

Build spatial, task, and interaction knowledge from embodied question answering and agent data.

Domain specialization

Learn manipulation and navigation decisions from verified successes and successful recoveries.

Multi-expert consolidation

Bring both experts into one student with multi-teacher on-policy distillation (MOPD).

Recursive self-improvement

The RSI data flywheel turns execution feedback and verified trajectories into model updates.

RSI data flywheel: interaction, verified trajectory pools, expert updates and MOPD consolidation

Collected successes, failures, and inefficient attempts reveal execution challenges. Verified successes and recovery trajectories update the domain experts, which are consolidated into the shared brain for the next collection round.

Real-world execution

Physical robot evaluation includes block stacking, apple placement, and drawer opening.

Physical robot execution: stacking a yellow cube on a red cube and placing an apple in a purple bowl, with structured decisions and tool calls

Citation

BibTeX
@techreport{omnibrain0_2026,
  title = {OmniBrain-0: A Unified Agentic Model for Embodied Intelligence},
  author = {{Xiaomi Embodied Intelligence Team}
            and {Tsinghua University} and {University of Macau}},
  year = {2026}
}