AEROS: A Single-Agent Operating Architecture with Embodied Capability Modules

Xue Qin; Simin Luan; John See; Cong Yang; Zhijun Li

cs.ROApr 8, 2026

AEROS: A Single-Agent Operating Architecture with Embodied Capability Modules

Authors: Xue Qin, Simin Luan, John See, Cong Yang, Zhijun Li

Organizations: School of Software, Harbin Institute of Technology, Harbin, China · School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China · School of Mathematical and Computer Sciences, Heriot-Watt University, Malaysia Campus, Putrajaya, 62200, Malaysia · School of Future Science and Engineering, Soochow University, Suzhou, China

Abstract

Robotic systems lack a principled abstraction for organizing intelligence, capabilities, and execution in a unified manner. Existing approaches either couple skills within monolithic architectures or decompose functionality into loosely coordinated modules or multiple agents, often without a coherent model of identity and control authority. We argue that a robot should be modeled as a single persistent intelligent subject whose capabilities are extended through installable packages. We formalize this view as AEROS (Agent Execution Runtime Operating System), in which each robot corresponds to one persistent agent and capabilities are provided through Embodied Capability Modules (ECMs). Each ECM encapsulates executable skills, models, and tools, while execution constraints and safety guarantees are enforced by a policy-separated runtime. This separation enables modular extensibility, composable capability execution, and consistent system-level safety. We evaluate a reference implementation in PyBullet simulation with a Franka Panda 7-DOF manipulator across eight experiments covering re-planning, failure recovery, policy enforcement, baseline comparison, cross-task generality, ECM hot-swapping, ablation, and failure boundary analysis. Over 100 randomized trials per condition, AEROS achieves 100% task success across three tasks versus baselines (BehaviorTree.CPP-style and ProgPrompt-style at 92--93%, flat pipeline at 67--73%), the policy layer blocks all invalid actions with zero false acceptances, runtime benefits generalize across tasks without task-specific tuning, and ECMs load at runtime with 100% post-swap success.

Figures & tables

Panel of Figure 1 — Figure 1: AEROS enables a single persistent agent to operate across simulated (a) and physical (b) platforms via installable Embodied Capability Modules. (c) Compared with published-architecture baselines across three tasks, AEROS achieves 100% task success ( $n{=}100$ , 95% Wilson CI).

	ROS 2	BT	LLM	Ours
Agent identity	—	—	$\circ$	$\bullet$
Capability pkg.	$\circ$	$\bullet$	—	$\bullet$
Persist. memory	—	—	$\circ$	$\bullet$
Policy safety	—	—	—	$\bullet$
Re-planning	—	$\circ$	$\bullet$	$\bullet$
Formal model	—	—	—	$\bullet$

Table 1: Qualitative comparison (

\bullet

= supported,

\circ

= partial, — = not supported).

Panel of Figure 2 — Figure 2: Single-agent robotic operating architecture. The persistent agent ( $A$ ) is the sole decision-making entity, invoking capabilities from Embodied Capability Modules ( $\mathcal{E}$ ). The runtime enforces policy constraints ( $\Pi$ ) and mediates all hardware interaction. Unlike multi-agent architectures, this design maintains a single intelligent subject with all capabilities provided as installable packages.

Panel of Figure 3 — Figure 3: Structure and lifecycle of an Embodied Capability Module (ECM). Left: internal schema showing skills, models, metadata, and permission declarations. Right: runtime lifecycle from installation to removal. The reactivation loop between Activate and Deactivate reflects that ECMs can be suspended and resumed without reinstallation. Notably, ECMs contain no agent-level constructs (identity, memory, planner), reinforcing that they are capability extensions, not independent subjects.

Panel of Figure 4 — Figure 4: Closed-loop execution model of the single persistent agent. Observation $o_{t}$ updates the agent state $s_{t}$ ; the planner $\mathcal{F}$ generates task $\tau_{t}$ ; the dispatcher selects skills $\sigma_{i}$ from installed ECMs $\mathcal{E}$ ; the runtime enforces policy $\Pi$ before execution. Environmental feedback $o_{t+1}$ closes the loop.

Method	Succ. (%, 95% CI)	Steps	Replan
Static	70.0 [60.5, 78.2]	3.7	1.0
Dynamic	100.0 [96.4, 100]	4.4	5.0

Table 2: Dynamic re-planning vs. static planning (100 trials).

Strategy	Succ. (%, 95% CI)	Steps	Recovery
None	49.0 [39.3, 58.8]	3.5	0.0
Retry	79.0 [70.0, 86.1]	4.3	0.0
Retry + Recov.	100.0 [96.4, 100]	5.2	0.3

Table 3: Impact of retry and recovery (100 trials, 50% failure rate).

Setting	Block (%)	F. Accept (%)	F. Reject (%)
Policy Disabled	0.0	100.0	0.0
Policy Enabled	100.0	0.0	0.0

Table 4: Policy enforcement (1800 permission checks).

Architecture	Dumpling (%)	CleanTbl (%)	FetchObj (%)	Mean
Flat Pipeline	73.0 [63.3, 80.8]	57.0 [47.2, 66.4]	73.0 [63.3, 80.8]	67.7
BT.CPP (retry=3)	95.0 [88.8, 97.8]	90.0 [82.4, 94.8]	93.0 [86.1, 96.7]	92.7
ProgPrompt (replan=3)	95.0 [88.8, 97.8]	90.0 [82.4, 94.8]	93.0 [86.1, 96.7]	92.7
AEROS (full)	100.0 [96.3, 100]	100.0 [96.3, 100]	100.0 [96.3, 100]	100.0

Table 5: Experiment 4 — Published-baseline comparison (4 architectures

\times

3 tasks,

n=100

per condition). CI = 95% Wilson score interval. Fisher’s exact test (AEROS vs. BT.CPP): dumpling

p=0.030

, clean_table

p<0.001

, fetch_object

p=0.007

Task	Static (%)	95% CI	Dynamic (%)	95% CI
Dumpling	72.0	[62.5, 79.9]	100.0	[96.3, 100]
CleanTbl	51.0	[41.3, 60.6]	100.0	[96.3, 100]
FetchObj	63.0	[53.2, 71.8]	100.0	[96.3, 100]

Table 6: Experiment 5 — Cross-task generality: static vs. dynamic re-planning on three tasks (

n=100

each). CI = 95% Wilson score interval.

Metric	Result (%, 95% CI)
ECM swap success	100.0 [96.3, 100]
Post-swap task success	100.0 [96.3, 100]
Overall (both tasks)	100.0 [96.3, 100]

Table 7: Experiment 6 — Runtime ECM hot-swapping (

n=100

). CI = 95% Wilson score interval.

Variant	Dumpling (%)	CleanTbl (%)	FetchObj (%)	Mean
AEROS (full)	100.0 [96.3, 100]	100.0 [96.3, 100]	100.0 [96.3, 100]	100.0
AEROS-no-policy	100.0 [96.3, 100]	100.0 [96.3, 100]	100.0 [96.3, 100]	100.0
AEROS-static-plan	97.0 [91.5, 99.0]	86.0 [77.9, 91.5]	100.0 [96.3, 100]	94.3
AEROS-no-recovery	97.0 [91.5, 99.0]	86.0 [77.9, 91.5]	92.0 [85.0, 95.9]	91.7

Table 8: Experiment 7 — Ablation study: AEROS variants (

n=100

per condition). CI = 95% Wilson score interval.

Figure 5: Experiment 8 — Failure boundary: mean task success rate vs. skill failure probability, averaged across three tasks (

n=100

per condition per task). AEROS maintains

>

97% success up to

p_{\text{fail}}=80\%

and degrades gracefully at 90%, whereas published baselines fall below 90% success at

p_{\text{fail}}=50\%

Design Decision	BT.CPP Original Colledanchise and Ögren (2018)	Our Reimplementation	Divergence
Execution model	Tick-based traversal	Tick-based traversal	None
Fallback strategy	Sequence + Fallback nodes	Sequence + Retry( $k$ =3) decorator	Equivalent retry semantics
Re-planning	Not supported (static tree)	Not supported (static tree)	None
Recovery actions	Via explicit Fallback subtrees	No explicit recovery subtrees	Simplified
Visualization	Groot GUI monitoring	Not included	Non-functional
Design Decision	ProgPrompt Original Singh et al. (2023)	Our Reimplementation	Divergence

Table 9: Baseline faithfulness: design decisions preserved and divergences from the original published systems.

Modular Robotics

Modular robotics studies robots built from interchangeable or self-reconfiguring parts that adapt their shape, capabilities, and tasks.

See all 46 papers

Evolving Skill Modules under a Fixed Planner: Versioning, Rollback, and Runtime Governance for Long-Lived Robot Systems
Apr 9, 2026Xue Qin, Simin Luan, Cong Yang +1Robotic Policy Evaluation Robot Safety
School of Software, Harbin Institute of Technology, Harbin, China · School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China · School of Future Science and Engineering, Soochow University, Suzhou, China
ETA: A New Agentic Paradigm for Embodied Tasks
Aug 4, 2026Yitong Chen, Zezheng Huai, Sixian Li +7Tool-Using Agents Robot Task Planning
1Shanghai Innovation Institute · 2Fudan University · 3Jilin University +2
EMERGE-Policy: A Robot Mind Emerges Beyond a Single Policy
Aug 30, 2026Zhirui Fang, Qingchi Yu, Ziyang Chen +18Cognitive Architectures for AI Agents Robot Policy Learning
Tsinghua University · Nanjing University of Science and Technology · Xi’an Jiaotong University +3

Apr 9, 2026J/K move · Enter open · S save

Evolving Skill Modules under a Fixed Planner: Versioning, Rollback, and Runtime Governance for Long-Lived Robot Systems

Xue Qin, Simin Luan, Cong Yang, Zhijun Li

School of Software, Harbin Institute of Technology, Harbin, China · School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China · School of Future Science and Engineering, Soochow University, Suzhou, China

Robotic Policy Evaluation Robot Safety

Open paper AskPDF arXiv

Robots deployed for long periods keep improving their skills, and each update changes a released system. We treat this as a software-lifecycle problem: a fixed decision layer dispatches versioned skill modules and a runtime layer was built to screen each action. On six robosuite tasks we report three negative results and two measurements. First, peak task success is unstable across random seeds (within one method it spans 23.3 to 73.3%), so single-run peaks cannot rank these methods. Second, the system's four modules are whole-task policies with different labels, rotated on a clock, not the phase decomposition its documentation describes. At a matched budget one such policy holds the geometry at the final step in 0.734 of episodes reaching it, averaged over seeds, against 0.023 for the rotation, with no seed overlap at four seeds per arm (exact p=0.029). An intervention isolates why: restoring the termination condition the clock replaced raises retention on every seed. Third, our shield cut violations 98 to 100% on five single-arm tasks (34.9% on the sixth) by discarding whole actions, leaving success at zero: its acceptance criterion omitted completions, so a shield that stopped the robot scored perfectly. What survives is release machinery: a promotion gate kept all twelve injected regressions out, a rate its calibration nearly guarantees, at a 22.5% clean-candidate rejection cost; a dip detector caught nine of twelve, missing all three on one seed.

Abstract

Modular Robotics

Explore similar work

Evolving Skill Modules under a Fixed Planner: Versioning, Rollback, and Runtime Governance for Long-Lived Robot Systems

ETA: A New Agentic Paradigm for Embodied Tasks

EMERGE-Policy: A Robot Mind Emerges Beyond a Single Policy