AEROS

The runtime that makes embodied agents deployable.

One robot. One persistent agent. One governed runtime.
AEROS provides the missing runtime layer for embodied intelligence: execution control, policy enforcement, recovery, and auditability.

v0.10.0 released · 2,239 passing tests · identity-hash regression on every PR

Beyond demos. Toward governed deployment.

Sim G1 walking end-to-end through Runtime → ECM → Governance → Embodiment. Live overlay shows the actual data shape the bridge emits — identity hash, ECM dispatch, policy gate, watcher tick, audit chain block. See the full chain →

Models can plan. Robots can act.
But embodied agents still break at runtime.

Today's embodied AI stack is powerful but incomplete. Foundation models can reason. Robot systems can move. What is still missing is the runtime layer that allows agents to run continuously, safely, and under control in the real world.

Reasoning is not execution

A plan is not a runtime. Real agents need lifecycle control, state continuity, task supervision, retries, and structured recovery when the world changes.

Hardware diversity breaks portability

Humanoids, mobile manipulators, robotic arms, and service robots should not each require a completely different agent stack. Embodied systems need a common runtime abstraction above hardware-specific execution.

Deployment demands governance

In production environments, capability calls cannot run without checks. They need contracts, policy enforcement, monitoring, rollback paths, and full audit trails.

How AEROS works

AEROS turns embodied AI from one-off task execution into a governed runtime process.

1 User goal
2 Agent planning
AEROS takes over
3 Capability admission
4 Contract validation
5 Policy check
6 Execution
7 Watcher monitoring
8 Recovery / rollback
9 Audit log

Every call is a governed event

Instead of blindly forwarding plans into robot action, AEROS validates what can run, under what conditions, with what monitoring, and with what fallback behavior if execution fails.

Steps 1–2 happen outside AEROS — any agent or planner can propose actions. From step 3 onward, AEROS takes control: checking admission, validating contracts, enforcing policies, supervising execution, and recording everything.

Nothing reaches hardware without passing through the governed runtime.

Not another Robot OS

ROS 2 focuses on communication
AEROS focuses on governed agent runtime
Workflow tools focus on orchestration
AEROS focuses on persistent execution under runtime constraints
Model stacks focus on reasoning
AEROS focuses on admission, supervision, recovery, and policy control

Linux gave machines a system layer.

Kubernetes gave distributed applications a control layer.

AEROS brings governed runtime control to embodied agents.

Built for real embodied deployment scenarios

Humanoids

Run persistent embodied agents with governed skill execution, lifecycle control, and safe capability evolution.

View benchmarks →

Mobile Manipulation

Coordinate navigation, perception, and manipulation under a unified runtime with monitoring and recovery.

View benchmarks →

Industrial & Enterprise

Add policy enforcement, rollout control, and auditability to production robot-agent workflows.

View benchmarks →

Cross-Embodiment

Reuse runtime logic across robot forms without rebuilding the full agent stack for every embodiment.

View benchmarks →

A five-layer architecture for embodied runtime

Layer 5: Agent Application logic & decision-making
Layer 4: Governance Policy enforcement & audit
Layer 3: ECM Package Capability modules & contracts
Layer 2: Runtime Lifecycle, scheduling, execution
Layer 1: Embodiment Hardware abstraction & sensors

Governance as a
first-class primitive

Traditional robotic middleware treats governance as an afterthought — a monitoring layer bolted on after deployment. AEROS makes governance a runtime primitive. Every capability call passes through the governance engine before reaching hardware.

The result: agents that are safe by construction, not by hope. Capability evolution that is controlled, not chaotic. Audit trails that are complete, not sampled.

Read the Architecture

Together, these layers separate intelligence from execution discipline, and execution discipline from hardware dependence.

What's shipped, with versions

v0.9.0 is the public baseline. v0.10.0 is in flight on main: identity, persona, memory, dreaming, humanoid — five subsystems, one frozen identity hash.

v0.9.0

Runtime Core

Lifecycle, scheduling, execution loop, fault handler. Apache 2.0.

v0.9.0

ECM Package System

6-dim typed contracts, v1/v2 dispatcher, shadow executor. Apache 2.0.

v0.9.0

Governance Layer

Policy DSL, watchers, 4-level recovery, Ed25519 audit chain.

v0.10.0

Frozen Identity

7-field IdentityManifest; identity_hash byte-equal across upgrades. HR-1 regression in CI.

v0.10.0

Adaptive Persona

Asymmetric trust matrix: tighten = auto-emit, relax = operator-signed. 20/day cap with floor-clamp metric.

v0.10.0

Layer 2 Semantic Memory

4 fact kinds, deterministic SQL aggregation, outbox-pattern emit durability. Relapse-safe.

v0.10.0

Active Dreaming

Offline reflection runner + scheduler. Placeholder LLM guard refuses dream cycles when no real model is wired.

v0.10.0

Unitree G1 Humanoid

ONNX bipedal locomotion (AGPL-clean). Walks pelvis ≥0.4 m / 1.5 s. Live integration test.

Research-backed, system-driven

AEROS is grounded in a broader research program around embodied runtime systems, governance, capability contracts, and deployment benchmarks. The goal is not to publish ideas in isolation, but to define a new systems layer for embodied intelligence.

Explore AEROS

For developers

Read the docs, inspect the runtime core, and explore the architecture.

For partners

See how AEROS enables governed deployment for embodied agents in real-world systems.

For researchers

Follow the research roadmap for runtime, governance, and embodied systems infrastructure.

The governed runtime for embodied agents.