Changelog

What shipped, when, and why.

The public repository ships as a consolidated open-source baseline; the entries below record the milestones that led to it.

v0.15.0 Open source

The full runtime goes public

2026-07 · github.com/s20sc/aeros-core

Everything ships: governance, ECM system, identity, persona, memory, dreaming, evolution, fleet, marketplace, migration, perception, navigation, MuJoCo sim backends, MCP server, provider SDK, OpenClaw plugin. No held-back "enterprise edition" — the kernel and the governance layer are the same code, in the same repository.

  • Dual license — AGPL-3.0 for the kernel; Apache-2.0 for the integration surfaces (planner interfaces, provider SDK, MCP server, contract schemas, ECM library, OpenClaw plugin). Commercial licensing available for deployments that cannot adopt the AGPL.
  • Test baseline — 2,483 passing tests in the public repository, 0 failing.
  • Docker path — docker compose --profile bridge-sim up --build brings up the governed bridge + MuJoCo sim on a fresh machine.
What's held back, and why

The benchmark suite + leaderboard, narrated demo drivers, and the bridge-side executors for perception/navigation follow later — on maturity grounds, not licensing ones. Each ships when it clears the bar the in-repo modules already meet.

View the repository Quick start
v0.11.0 Hardware track continuing

Hardware adapters & EmbodiedGovBench v2

First batch landed 2026-05-09 · end-to-end real-hardware demo still pending

v0.10.0 ships the governance core. v0.11.0 takes that core out in two directions: real-hardware adapters (Unitree G1 + Franka) and a public, multi-LLM benchmark with leaderboard.

  • Real Unitree G1 adapter — RealUnitreeG1Robot implements the Robot ABC against Unitree SDK 2 RPC. identity_hash byte-equal across g1_sim ↔ g1_real swap.
  • Real Franka adapter — RealFrankaRobot against the franka_ros2 driver, with workspace-bounds + speed-factor defense in depth.
  • Humanoid scenario suite — 4 G1 scenarios for NavigateECM + GraspAndPlaceECM, including an enforcement check that operator-relax without signature must be refused.
  • Multi-LLM provider harness — matrix runner across Claude / GLM / Gemini / local-onnx. --cheap offline mode for CI.
  • Leaderboard — regenerable, every row commit-pinned; public release rides with the benchmark suite.
  • Experiment reproducibility — three core experiment claims (audit-first rollback, byte-identical replay, grounded planner) now run as pytest suites; existing scripts kept as thin wrappers.
  • Operator docs — benchmark + scenario guides covering full-matrix runs, env vars, cost estimates, scenario authoring.
What's deferred

The ROS 2 bridge process and the end-to-end real-hardware demo (G1 walks 1m + Franka grasps under live bridge) are deferred until a maintainer has both robots in physical reach. Those land as the v0.11.0 final tag along with the full hardware-track release notes.

v0.10.0 Released

Adaptive governance — identity, persona, memory, dreaming

Released 2026-05-09 · release notes

v0.9.0 shipped the governance bones: policy DSL, watchers, audit chain, recovery orchestrator. v0.10.0 makes governance adaptive — the robot can learn from its own episodic stream, evolve a persona, and reflect offline, all without flipping its identity hash.

  • Frozen Identity — IdentityManifest with seven hash inputs; identity_hash is byte-equal across upgrades, memory growth, and persona evolution. Identity-hash regression on every PR.
  • Adaptive Persona — PersonaAdaptiveEvent typed vocabulary; asymmetric trust matrix (tighten = auto, relax = operator-signed Ed25519); 20/day cap with floor-clamp metric.
  • Layer 2 Semantic Memory — four fact kinds (skill_success_rate, object_property, zone_risk, interaction_pattern) with deterministic SQL aggregation, 100-event / 6h triggers, outbox-pattern emit durability, fail-closed timeouts.
  • Active Dreaming — offline reflection runner + DreamScheduler. Placeholder LLM guard refuses dream cycles when no real model is wired.
  • Unitree G1 Locomotion — ONNX bipedal locomotion policy with a dependency stack that passes the third-party license gate; SimulatedUnitreeG1Robot.apply_twist walks pelvis ≥0.4 m / 1.5 s; live integration test wires NavigateECM to the real policy.
  • Test baseline — ~2,239 passing on main, 3 skipped, 0 failing (excluding known X11/EGL render flake).
Why this shape

The wedge: governance between LLM and actuator. v0.10.0 is what that means in code — identity is the substrate, persona is the policy, memory is the evidence, dreaming is the offline catch-up, humanoid is the proof-of-life. Five subsystems, one invariant: identity stays frozen through all of it.

Read the architecture
v0.9.0 Released

Governance bones — policy, watchers, audit, recovery

2026-04 · internal milestone

The governance core that makes everything else possible. Every capability call passes through admission, contract validation, policy check, watcher monitoring, recovery orchestration, and audit logging.

  • Runtime Core — lifecycle, scheduling, execution loop, fault handler.
  • ECM Package System — six-dimensional contract, v1/v2 dispatcher pattern, shadow executor.
  • Governance Layer — policy DSL, watcher system, four-level recovery (log / alert / intervene / halt), Ed25519 audit chain.

Earlier releases

  • v0.7.0 Pre-v0.9 architecture iterations
  • v0.6.0 ECM contract refinement
  • v0.5.0 Baseline

Earlier iterations were internal; the public repository ships as a consolidated v0.15.0 baseline.

Want to see the numbers behind the claims?

Three measured guarantees: rollback success, replay determinism, planner action quality.

View benchmarks →