↳ Release history

Every release, in the open.

The reference packages ship as one lockstep family. The latest is v0.4.0 (pip install urml-validator); the normative spec documents version per-artifact and are at v0.2.0. The full history is below, generated from the canonical CHANGELOG.md in the core repo.

CHANGELOG on GitHub ↗ Read the spec

Latest release

v0.4.0 2026-08-09

5 released versions since 2026-05-22. Spec docs at v0.2.0; the v1.0 freeze is targeted later, see governance.

On main since v0.4.0 Not yet released

Work on main since 0.4.0, not yet cut into a release. The packages on main still carry version 0.4.0. PyPI serves 0.4.0 of urml-validator, urml-llm-bridge, and four runtimes (urml-ardupilot-runtime, urml-autosar-runtime, urml-chrono-runtime, urml-cobot-runtime), uploaded 2026-09-10, and those builds predate most of the work below. The other packages, including urml-ros2-runtime (the mock runtime), urml-px4-runtime and urml-mcp-server, install from source; install every URML package from the same source, because a runtime built from main needs the validator from main.

Added: ArduPilot reference runtime

·urml-ardupilot-runtime (ArduCopterAdapter): ArduCopter over MAVLink through pymavlink, no ROS (#720). On a Pixhawk-class board with ArduCopter 4.6.3, on USB with propellers off (2026-08-29), the bench-battery program ran to success. The bench-hop program passed the validator and the autopilot's own pre-arm checks refused to arm; URML reported the autopilot's reason and stopped. No free flight on hardware is claimed.
·ArduCopter SITL (Copter-4.6.3, run locally): the flight-only conformance fixture, the site-photogrammetry example and the parcel-delivery example passed, 3 of 3 (#721). The run surfaced one defect: ArduCopter 4.6 rejects the MAVLink winch deliver and retract actions, so the adapter uses relative-length control. On 2026-08-30 the programs Gemini 2.5 Pro emitted through the LLM bridge for the two flight-test sentences flew unchanged on the same SITL (#724).
·Payload release through winch and latch (RFC-0684, #771).

Added: first real adapters for four more kinds of behavior

·Trajectory: urml-av-runtime ships AutowareAdapter, the first real plan_path / follow_trajectory adapter (RFC-0020, #774), wired into the MCP server's execute path (#775).
·Expression: ReachyMiniAdapter in the edu runtime, the first real expression adapter (RFC-0698, #777). Hermetic tests only; it has not run on a robot yet.
·Relative motion: GoPiGo3Adapter in the edu runtime, the first real drive / turn runtime (RFC-0630, #773).
·Arm: the Spot adapter gains grasp and release behind arm_attached, with a Spot-with-Arm manifest (#700). The capture primitive gains a camera selector (RFC-0699, #740).
·MicroduckAdapter for the Hugging Face duck over its JSON-RPC contract, with the velocity cap held to the policy's training range (#727, #728).

Added: rulebooks (RFC-0702)

·A rulebook pass for government and company rules (#756, #759). The bundled rulebook, the statically checkable subset of 14 CFR Part 107, applies to drone programs by default. --no-policy does not switch it off; --no-default-rulebooks does, with a warning. --rulebook PATH adds an organization or deployment rulebook. 13 conformance fixtures under conformance/fixtures/rulebook/; spec text in spec/layer-1-hal/rulebook.md.

Added: validation records and the registry

·--evidence-log PATH appends one validation record per verdict, accepted or refused, with content digests, at every entry point (#762). Opt-in and local.
·registry/: a public registry of robots and runtimes validated with URML, submitted by pull request and checked by python -m urml_conformance.registry check (#764). Three entries: ArduCopter on a Pixhawk-class board, the GoPiGo3 example adapter, and PX4 in SITL (#766). A compatibility claim covers a whole profile (#767). Rendered at https://urml.dev/registry.

Added: urml bench and clarify mode

·urml bench measures how well a model speaks URML, with starter corpora, adversarial corpora, a scripted worst-case striker, and a gate table per row (#743, #751).
·Clarify mode: with --clarify, urml translate and urml run let the model ask one clarifying question instead of guessing. Default off, experimental. Its Spec RFC (RFC-0700) is proposed and not yet merged (#746).

Added: safety model and the goalkeeper

·docs/safety/safety-model.md and docs/safety/envelope-coverage.md: what the validator checks for each primitive, and what it cannot see (#752, #758).
·examples/goalkeeper/: a scripted compromised model against three robots; every attack refused, no command sent (#754). Worst-case striker rows: 21 of 40 attacks got through before the fixes, 0 of 40 after, with 12 of 12 safe controls still accepted (#755).
·A goal-line conformance lane: every rejected fixture is refused with zero adapter calls (#750).
·The MCP server pins manifest, envelope, profiles and policy per server (#749).
·A STREL compile target for monitorable envelopes; RFC-0667 and RFC-0668 move to Implemented (#760). Dynamic-target grasping (RFC-0671, #753).

Added: examples

·A learned ExecuTorch policy behind the runtime shield (#688), a Decart Oasis world-model rehearsal (#680), Unitree G1 whole-body (#690), URDF to manifest (#678), MTConnect MOVE_MATERIAL lowered to a UR arm (#697), a physical-AI safety-eval harness over Anthropic's Model Hardware Standard (#705), and Times Square geocoding with NYC and FAA rules (#722).

Fixed: the PX4 adapter flies the program

·The adapter treated every COMMAND_ACK as success, so a SITL test could pass while the vehicle stayed on the ground. Each flight primitive is now confirmed from PX4 telemetry (#765). The adapter sends a ground-station heartbeat, and a refused command carries PX4's own reason (#769). Recorded SITL runs, local and not in CI: reference/px4-runtime/tests/integration/sitl-runs/.
·scan no longer reports success with made-up results, and grasp and capture calls no longer raise TypeErrors, across the reference adapters (#772).

Added: ROS 2 scan through Nav2

·RclpyAdapter runs a scan step as a real scan plan through Nav2 (#783). Run locally against Nav2's loopback simulator (real planner, controller and behavior tree over a kinematic robot, no physics), not in CI: one validated scan step drove Nav2 to all nine waypoints of a serpentine over a 2 m square, a listen-only witness saw the robot within 0.23 m of every waypoint, and a goal off the map came back as goal_aborted instead of a success. Record: reference/ros2-runtime/tests/integration/nav2-runs/2026-09-29-nav2-loopback-scan.md.

Removed: the Moltbook agent

·RFC-0640, its ledger, and tools/moltbook/ are gone (#781). The channel drew no response in seven weeks. The venue-neutral agent guide, docs/integrations/urml-for-ai-agents.md, stays. RFC numbering keeps a gap at 0640.

Changed: urml translate and urml run carry the request into description

·The emitted program's existing optional top-level description now holds the sentence (or the spoken transcript) it was translated from, placed right after profile. A description the model wrote itself is kept. Runtimes can print it next to the lowered calls; the GoPiGo3 example does, as prompted by:. No schema change (Discussion #597).

Fixed: GoPiGo3 example

·wait is honored. The adapter takes a wait clock (default time.sleep) and the runner gates it exactly like speech: --execute holds the robot for the declared duration, a dry run records time.sleep(N) in the plan and never blocks. The report line now shows the duration (Discussion #600, Issue #592).
·PR and issue templates link CONTRIBUTING.md, CLAUDE.md, and MANIFESTO.md by absolute URL. Relative links do not resolve from a PR or issue page (Discussion #604).

v0.4.0 2026-08-09 Latest

Minor release. Everything is backward-compatible: a v0.3.0 manifest, envelope, and program validate and execute unchanged. All 20 Python packages move to 0.4.0 in lockstep, and 0.4.0 is the first version of the family actually uploaded to PyPI (0.2.0 and 0.3.0 were version alignments on main; the index still served 0.1.0).

Added — Speech front-end (RFC-0670)

·urml translate --audio request.wav and urml run --audio request.wav: speak the request instead of typing it. A SpeechProvider protocol mirrors the LLM provider contract, with adapters for whisper.cpp's whisper-server (new [whisper_cpp] extra, on-device, the default), OpenAI-compatible transcription endpoints (key not required with a base URL), and a hermetic echo provider. The transcript is echoed to stderr and feeds the unchanged translate path, so the validator boundary is untouched. --speech-language hints the language for multilingual requests.

Added — First-class on-prem LLM providers in the CLI

·urml translate and urml run gain --provider ollama and --provider llama_cpp, wiring the RFC-0021 on-device adapters (shipped and tested since v0.2) into the CLI. A new --base-url flag overrides the local server endpoint for both, and points --provider openai at any OpenAI-compatible server (LM Studio, vLLM).
·OPENAI_API_KEY is no longer required when a base URL is set via --base-url or OPENAI_BASE_URL: local servers ignore the key, so the CLI passes a placeholder. Cloud usage without a base URL is unchanged.
·A dead local server now surfaces as one actionable line naming the endpoint and the fix, instead of a raw httpx traceback.

Notes

·The PyPI publish (TestPyPI rehearsal, then the real index), the v0.4.0 git tag, the GitHub release, and the README install-line flip are executed by the maintainer after this lands, per RELEASING.md.
·New tools/scripts/bump_version.py: the lockstep bump is now scripted and checkable (--check), replacing the hand-edited 40-file sweeps of 0.2.0 and 0.3.0.

v0.3.0 2026-07-23

Minor release, and the first to version the full package family in lockstep. Everything is backward-compatible: a v0.2.0 manifest, envelope, and program validate and execute unchanged. All 20 Python packages move to 0.3.0 in lockstep.

Added — Envelope enforcement (RFC-0667)

·Normative satisfaction semantics for the RFC-0382 temporal-logic core over finite timed traces: completed-trace offline evaluation, three-valued online verdicts (satisfied / violated / pending), strong truncation semantics, and the severity mapping (info records, warning audits, critical vetoes).
·urml_validator.monitor: the reference trace evaluator (evaluate_trace, OnlineMonitor, compile_envelope_monitors). Static envelope caps (max_velocity, max_grip_force_n, altitude) now derive to implicit critical always properties, so one evaluator covers declared and derived properties alike.
·urml_ros2_runtime.shield: the runtime shield. Shield gates each dispatch and observes telemetry; ShieldedAdapter wraps any substrate adapter so any driver, a URML runtime or an external policy, passes through envelope checks. New optional TelemetryAdapter side-protocol; the frozen RFC-0014 adapter contract is untouched. External runtime-verification backends stay first-class through compile_to_stl.

Added — Rehearsal gate and the urml run verb (RFC-0668)

·urml run "<sentence>": translate, validate, optionally rehearse in simulation, then execute, in one verb. --rehearse is also available on urml execute.
·Rehearsal backends: a hermetic synthetic-kinematics backend (CI floor, honest about being kinematics rather than physics) and a MuJoCo per-tick recording adapter with a declared SignalMap. A critical violation in the rehearsed trace blocks execution before any real adapter is constructed.
·The README hero now shows the gate live: the same sentence first blocked (envelope speed cap violated under the default motion assumption), then passing under the deployment's declared profile.

Added — URML-native model pipeline (RFC-0666)

·New package urml-model (reference/model): mines conformance gold programs and the curated few-shot library, back-translates programs to utterances (deterministic templates in CI, any LLMProvider at scale), filters every record through the validator oracle, and exports SFT/DPO training files. Ships the operator-run LoRA training recipe. No weights, no vendor coupling, nothing leaves the repo checkout.

Added — Agent surface

·urml-mcp-server (reference/mcp-server): URML's validate-before-actuate loop as Model Context Protocol tools for any MCP-capable agent. The server never calls an LLM; the calling agent emits URML, the server checks and runs it.
·An Agent Skill (.github/skills/urml-robot-intent) with a hermetic no-key, no-robot path.

Added — Specification and manifests

·Per-capability evidence traceability (RFC-0631): capability claims carry derived / verified / declared / inferred evidence strength, with an opt-in policy to require the stronger grades.
·Whole-body kinematic structure and stability limits for legged and humanoid platforms (RFC-0384); whole-body and bimanual manipulation (RFC-0010).
·Realtime cyclic-timing manifest block (RFC-0016) and the zero-copy IPC sub-substrate declaration (RFC-0385).
·drive radius parameterization (RFC-0665): a radius plus sweep angle pass through without arc-length arithmetic, closing the small-model orbit-2x failure observed in the field.

Fixed

·GoPiGo3 example runtime converts URML's FLU (+CCW) frame to the platform's RFD (+CW) frame at the hardware boundary (#591, #598, PR #602). URML's first merged external code contribution, found and fixed on real hardware by @slowrunner.

v0.2.0 2026-06-02

Minor release. The first spec-bumping release since v0.1.0: it transcribes the accepted multi-robot fleet RFCs into the normative layer specs, so the specification stops lagging its own accepted decisions. Layers 1, 3, and 4 advance to v0.2.0; Layer 2 is unchanged (still 20 normative primitives). All additions are backward-compatible — a v0.1.0 manifest and a single-robot program validate unchanged.

Added — Specification

·Layer 1 (HAL) → v0.2.0. Three additive, optional items: a frame transform (SE(3) pose in the parent frame), a mobility.clearance operational volume (lateral footprint + vertical altitude/depth band), and the new §6 fleet roster (binds N per-robot manifests by handle; world_frame, shared_frames, per-member anchors) plus the validate_fleet cross-robot checks. RFCs 0286, 0290, 0291. manifest_version stays "0.1".
·Layer 3 (behavior composition) → v0.2.0. Two additive composition nodes: on (scope a subtree to a fleet member) and barrier (synchronize members at a rendezvous). RFC 0286.
·Layer 4 (NL grammar) → v0.2.0. Roster-aware fleet (multi-member) prompt assembly and a fleet revision loop (validate_fleet), emitting on/barrier programs. RFC 0286.
·Sensor schema v0.2 reconciled. The RFC-0039 sensor capability fields (beam_count, channels, time_sync_methods, rate_hz_max, point_cloud type), already shipped in the Layer-1 doc since v0.1.1, are now carried normatively at v0.2.0. RFC 0039.

Added — Reference implementation

·Multi-robot fleet vertical slice. validate_fleet (four cross-robot checks plus shared-frame/anchor validation), a concurrent FleetRuntime, the courier-to-arm handoff demo, and a fleet conformance lane (conformance/fixtures/fleet/, 14 cases covering ground/air/water separation and the engaged-partners choreography). RFCs 0286, 0290, 0291.
·urml-edu-runtime: fourth platform. RoboticalMartyAdapter graduated to production (real-martypy API-surface CI gate), joining VEX V5, LEGO SPIKE, and Thymio.

Notes

·All 17 in-repo packages (5 published on PyPI + 12 in-repo runtimes) bump 0.1.1 → 0.2.0 in lockstep.
·This is strategic cross-robot deconfliction: static rejection of plans that would put two robots in one volume at one time, before execution. It is not a live collision-avoidance loop and not a building-scale traffic manager (that is Open-RMF's domain; URML composes above it).
·The PyPI publish, the v0.2.0 git tag, and the README/Tutorial-01 install-instruction flip follow RELEASING.md and are gated on the maintainer.

v0.1.1 2026-05-26

Patch release. Phase 1 maintenance since v0.1.0: the first outreach-feedback-driven adapter, two committed homepage diagrams with byte-exact guard tests, and the PyPI packaging fix that should have shipped earlier. No spec change (RFC-0039 sensor schema v0.2 remains in Draft); no breaking change.

Added — Reference implementation

·urml-edu-runtime: RoboticalMartyAdapter. New adapter for the Robotical Marty platform, scaffolded in response to early outreach feedback. Conformance fixtures stay hermetic against MockROSAdapter. RFC 0073; commit d516535.

Added — Docs & assets

·Architecture-stack diagram. Committed, deterministic SVG of the five-layer URML stack at docs/assets/architecture-stack.svg, generated by tools/scripts/gen_architecture_svg.py and pinned to docs/architecture.md by reference/validator/tests/test_architecture_svg.py. Same pure-stdlib, byte-exact, CSS-@keyframes-only discipline as the README hero. PR #129.
·One-intent-many-bodies portability demo. Committed SVG showing the same URML program executing across multiple substrate adapters, byte-asserted against a live hermetic run. PR #132.

Fixed

·urml-validator: PyPI upload. Redundant force-include in the package build config was breaking PyPI uploads. Removed. Commit 1d631b4.

Notes

·All 17 in-repo packages (5 published on PyPI + 12 in-repo runtimes) bump 0.1.0 → 0.1.1 to preserve the alignment policy from v0.1.0. None of the 12 in-repo runtimes ship to PyPI in v0.1.1 either; that follow-up is per-package and tracked separately.
·Layer-1 sensor schema is normative at v0.1.0 still; the v0.2 iteration (RFC-0039) lands in a future spec-bumping release once Accepted.
·Phase 1 cosmetics on urml.dev shipped in the website repo (separate from this changelog): why-now hook, architecture SVG, audience-segmented CTAs, Phase 0 → Phase 1 chrome refresh (urml-website#11, #12).

v0.1.0 2026-05-22

First public release. URML moves from Phase 0 (solo-author working in public) to Phase 1: pip install urml-validator (and four sibling packages) ships, external contributions open per CONTRIBUTING.md. The strategic posture, the Core Commitment, and the substrate-neutrality contract are unchanged from the manifesto; what changes is the gate.

Added — Specification

·Layer 1 (HAL). Capability manifest with provenance, safety envelope, and connectivity / link-loss as a validated safety contract. Normative at v0.1.0. RFCs 0001, 0002, 0006, 0009, 0014.
·Layer 2 (intent primitives). 20 normative primitives: 12 core (move_to, dock, hover, wait, wait_for, grasp, release, detect, scan, measure, capture, report) plus 8 profile-scoped (speak/listen home; take_off/land/return_to_home drone; pick_from/place_at/swap_tool industrial). RFCs 0002, 0013.
·Layer 3 (behavior composition). sequence / branch / parallel / retry + on_error (abort_and_report / continue / retry). Normative at v0.1.0.
·Layer 4 (natural-language bridge). Provider-agnostic prompt contract, validator-feedback revision loop with policy-error short-circuit, schema-derived GBNF + GGUF contract for on-device models. RFC 0021.
·Profiles. home, drone, industrial, educational, research, warehouse — each with its own README, default safety envelope, and conformance fixtures. RFCs 0011, 0012, 0022.
·Compliance policy enforcement. Provenance schema on the manifest, pluggable YAML policy DSL, bundled US-federal default policy aligning with NDAA §889 / FY26, FCC Covered List (effective 2025-12-23), Executive Order 14307, and the American Security Robotics Act once enacted. --no-policy opt-out. RFCs 0003, 0004, 0005.
·Substrate-conformance contract. What it means for a runtime to be URML-compatible, codified normatively. RFC 0014.

Added — Reference implementation (PyPI)

·urml-validator — five-pass static validator (argument typing → capability → safety envelope → variable bindings → compliance policy) and the urml CLI with seven subcommands: validate, execute, schema, translate, emit-prompt, init, conformance run.
·urml-ros2-runtime — MockROSAdapter (hermetic, zero-dep) plus RclpyAdapter against the live ROS 2 graph (Nav2, MoveIt 2, vision_msgs). The home/nav_patrol_positive conformance fixture is end-to-end verified ×3 on a TurtleBot 4 + Nav2 + Gazebo Ignition simulation via the gated gazebo-e2e CI job — green at the job level on three calibration runs; not a hardware-verification claim.
·urml-llm-bridge — provider-agnostic NL → URML bridge with Anthropic, OpenAI, hermetic Echo, plus on-device llama_cpp and ollama providers (RFC 0021). Schema-derived GBNF grammar; GGUF model contract; per-model conformance harness.
·urml-px4-runtime — PX4Adapter via pymavlink (no ROS dependency); CompositeAdapter routes a single URML program across a PX4 flight backend and a ROS 2 companion. Hermetic + gated SITL e2e.
·urml-conformance — 101 declarative YAML fixtures (auto-discovered) across 10 profiles, runnable via urml conformance run; the bring-your-own-adapter kit for runtime authors.

Added — Twelve further reference runtimes (in-repo, not on PyPI in v0.1.0)

·marine (BlueROV2 / ArduSub), industrial-arm (16 brand adapters: ABB / FANUC / KUKA / YASKAWA / UR / Franka / Kawasaki / Stäubli / Comau / Mitsubishi / Denso / Hyundai / Nachi / Epson / Omron / Hanwha), legged (Spot / ANYmal), humanoid (Digit), mobile (Husky / Jackal); the zero-ROS opcua (OPC UA Robotics) and cobot (8 native SDKs: UR / Franka / Doosan / Techman / Kinova / Mecademic / Neura / Kassow); mujoco and isaac (NVIDIA Sim/Lab, local RTX host) sims; embedded (micro:bit / Arduino serial); edu (VEX / LEGO SPIKE / Thymio); autosar (RFC 0019 scaffold). Autoware ships manifest + spec only pending RFC 0020. Each is one pip install -e <path> away from runnable; PyPI follow-up per package.

Added — Conformance & lighthouses

·101 conformance fixtures under conformance/fixtures/, hermetic against MockROSAdapter: home 18, drone 14, industrial 44, biped 5, quadruped 4, mobile 2, warehouse 8, marine 1, educational 4, research 1. Auto-discovered.
·Move #1 lighthouse program. Per-vendor request-for-comment RFCs to 16 Tier-1 vendors (RFCs 0023–0038): Yaskawa, Universal Robots, KUKA, Stäubli, Mitsubishi MELFA, FANUC, Kawasaki, Denso, SCHUNK, Ouster, SICK, Festo, Zivid, Hokuyo, OSRF / Gazebo Sim, ROS-Industrial Consortium. Per-vendor outreach state tracked in examples/lighthouses/outreach.yaml; the parameterized demo runner in examples/lighthouses/demo.py exercises any single vendor's conformance fixture hermetically.

Added — Tooling & developer experience

·make demo / make demo-run / make demo-record / make audit / make test. The flagship make demo-run is the universal-language pitch concretised: one English sentence becomes a validated URML program becomes an executed step-by-step trace, hermetic, any-OS, no API key, no robot. The animated SVG hero in the README is exactly what make demo-run prints (every line asserted in CI via reference/validator/tests/test_demo_svg.py).
·tools/scripts/refresh_audit.py — stdlib-only re-measurer for the front-page numbers; honest about rows the host cannot remeasure; does not auto-edit. make audit.

Added — Process & community

·RFC process (RFC 0001): Draft → Open → Accepted → Implemented with a 7-day Phase-0 comment window; 22 Spec RFCs and 16 Outreach RFCs filed and indexed (Kind column distinguishes the two; see docs/rfcs/README.md).
·Public GitHub Discussions open (Q&A, Ideas, Show-and-tell, Builders & Makers, General) per RFC 0008.
·Manufacturer go-to-market wedge per RFC 0007; the live manufacturer / federal-validation surface on urml.dev.
·US-federal regulatory alignment per RFC 0003. Strategic trade-offs accepted in writing.

Test surface (claims-audit, measured 2026-05-20 with partial 2026-05-22 re-run)

·765 passed + 28 gated-skipped across 16 reference packages.
·Live integration is gated CI (*-integration.yml, workflow_dispatch + weekly cron). First run of any live e2e is a calibration baseline, not a regression signal. Backing is recorded in docs/launch/claims-audit.md.

Honest scope (no overclaiming)

·No physical-hardware verification. Live e2e is simulated (Gazebo / PX4 SITL) or gated CI calibration. Where the README says "verified," it means the gated CI job was green; that is the load-bearing claim, not a hardware-on-a-bench claim.
·Compliance enforcement is not legal advice. The bundled default policy mirrors enacted US federal procurement law as of release date; deployers consult counsel. Disclaimers in docs/launch/claims-audit.md and the bundled policy file.
·Five canonical packages on PyPI in v0.1.0; the twelve further runtimes follow per-package as each is ready.
·One maintainer. Phase 1 opens external contributions; it does not assert contributors have arrived. The Status: one person line in GOVERNANCE.md remains true until it isn't.