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

Jul 16, 2026

Upload by Hazel

Embodied AI is shifting from standalone perception or action components to physical agents that understand the world, plan toward goals, act through robot bodies, monitor progress, and improve from experience. Prior approaches fragment this loop across end-to-end policies and tool-orchestrating agents, limiting shared robot representations and long-horizon execution assessment. ACE-Brain-0.5 unifies embodied robot intelligence into five coupled functions—spatial perception, decision making, embodied interaction, self-monitoring, and self-improvement—using an 8B closed-loop backbone and SSR+ training plus an external-state self-improvement framework.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages69

Reading Time2 h 54 min

Introduction

Related Work

General Multimodal Models, Spatial Reasoning, and Embodied Planning

End-to-End Manipulation Policy

Foundation Models for Embodied Navigation

Progress-based Reward Models

Self-Improving Robotic Agents

Overview

Model Architecture

Training Strategy: Scaffold, Specialize, Reconcile, and Reactivate (SSR+)

Self-Improving Framework

Experiments

Spatial Perception

Decision Making

Embodied Interaction

Self Monitoring

Self Improvement

Data and Benchmark

Training Data

Evaluation Benchmark

Conclusions

Contributions and Author List

Appendix

arXiv :2607 .04426v 1 [ cs .RO] 5 Jul 2026 ACE-Brain-0.5: A Unified Embodied Foundational Model for Physical Agentic AI ACE-Brain Team Please see Contributions and Author List for more author details. Abstract Embodied AI is moving from isolated perception or action modules toward physical agents that can understand the world, plan under goals, act through robot bodies, monitor whether their behavior is making progress, and improve from accumulated experience. Existing systems address different parts of this loop in isolation: end-to-end policies (vision-language-action or world-action) are effective at generating robot actions but often provide limited spatial reasoning, long-horizon planning, and execution assessment; robot-agent systems can orchestrate multiple tools or specialist models but do not learn a single shared robot representation. This fragmentation limits the development of general Physical Agentic AI. We present ACE-Brain-0.5, a unified embodied foundation model for Physical Agentic AI that organizes robot intelligence into five tightly coupled cognitive functions: spatial perception, decision making, embodied interaction, self-monitoring, and self-improvement. Built on ACE-Brain-0, which established spatial intelligence as a shared scaffold across heterogeneous robot platforms, ACE-Brain-0.5 extends an understanding-centric embodied model into a closed-loop embodied foundation model. A single 8B backbone directly instantiatesthe first four functions within a closed loop: it grounds objects and affordances, reasons over 3D and egocentric spatial relations, decomposes high-level instructions into executable subgoals, generates navigation and manipulation actions, and estimates execution progress for verification and recovery. To unify these heterogeneous capabilities without cross-task interference, we introduce SSR+, which extends Scaffold–Specialize–Reconcile with a lightweight Reactivate stage after task-vector merging. The fifth function, self-improvement, is realized through a companion framework that incrementally updates an external execution state, i.e., task schemas, spatial memory, and failurerecovery cases, from accumulated rollout experience, enabling deployment-time adaptation. Across more than fifteen benchmarks spanning spatial cognition, grounding, navigation, manipulation, and progress evaluation, ACE-Brain-0.5 improves over ACE-Brain-0 on 14 out of 18 spatial perception and grounding benchmarks, achieves competitive navigation and manipulation performance, and provides strong progress-estimation ability under both in-distribution and out-of-distribution settings. These results show that spatial perception, decision making, embodied interaction, self-monitoring, and self-improvement can be unified within a single robot foundation model, marking an early step toward general Physical Agentic AI. Date: July 7, 2026 Project Page: [https://ace-brain-team.github.io/ACE-Brain-0.5/](https://ace-brain-team.github.io/ACE-Brain-0.5/) Code: [https://github.com/ACE-BRAIN-Team/ACE-Brain-0.5](https://github.com/ACE-BRAIN-Team/ACE-Brain-0.5) Hugging Face: [https://huggingface.co/ACE-Brain/ACE-Brain-0.5-8B](https://huggingface.co/ACE-Brain/ACE-Brain-0.5-8B) Contents 1 Introduction ................................................ 3 2 Related Work ................................................ 5 2.1 General Multimodal Models, Spatial Reasoning, and Embodied Planning ........... 5 2.2 End-to-End Manipulation Policy .................................. 5 2.3 Foundation Models for Embodied Navigation ........................... 6 2.4 Progress-based Reward Models ................................... 6 2.5 Self-Improving Robotic Agents ................................... 7 3 Overview .................................................. 7 3.1 Model Architecture .......................................... 7 3.2 Training Strategy: Scaffold, Specialize, Reconcile, and Reactivate (SSR+)

FAQs

What problem does ACE-Brain-0.5 address in Physical Agentic AI?

It addresses the fragmentation of the perception–planning–action–monitoring–improvement loop, where existing systems cover parts of the pipeline in isolation and fail to learn a single shared embodied robot representation.

How does ACE-Brain-0.5 unify robot intelligence?

ACE-Brain-0.5 organizes robot cognition into five tightly coupled functions: spatial perception, decision making, embodied interaction, self-monitoring, and self-improvement, implemented via a closed-loop 8B backbone and supporting training components.

How is self-improvement implemented in ACE-Brain-0.5?

Self-improvement is handled by a companion framework that incrementally updates an external execution state—such as task schemas, spatial memory, and failure-recovery cases—using accumulated rollout experience for deployment-time adaptation.

Jul 16, 2026

Upload by Hazel