Lagrange v0.2.0: simulation core and AMR (release notes, draft)

v0.1 was the viewer, the protocol and a replay server. v0.2 is the simulator behind them: a multirate co-simulation core, MuJoCo physics, lidar and IMU sensors on the CPU and the GPU, a ROS 2 bridge with ros2_control, and a Nav2 demo, with the v0.1 viewer connecting unchanged.

Status: released on 2026-10-25 as v0.2.0, at the owner’s request, with acceptance criterion 1 still open (below).

Highlights

  • Core.

    • Units checked at wiring, with no silent conversions.

    • A signal bus and a multirate scheduler with holds and feedback inputs.

    • Substitutes for every domain; save, restore and state hashes.

    • Counter-based noise; a floating origin (a world 500 km out simulates like one at the origin).

    • Core services: wrench_sum, a run log, SceneQuery, bounded consumer rings.

    • A versioned C ABI for third-party modules (ABI 0.1, unstable until v1.0).

  • Physics.

    • MuJoCo 3.12, validated against closed-form mechanics.

    • The fixed-robot substitute.

    • A Jolt Physics spike, with its gaps listed for the full backend in v0.3.

  • Sensors.

    • IMU with white noise, bias, random walk, saturation and quantisation.

    • 2D and 3D lidar on Tier 0: exact casts in double, Embree-culled, bit-identical to testing every shape.

    • 2D and 3D lidar on Tier 1: a Slang compute kernel over a bounding-volume hierarchy.

    • Wheel encoders through ros2_control.

  • Server and tools.

    • lagrange_server speaks the v0.1 protocol to the viewer.

    • Simulation profiles are validated and applied.

    • MCAP recording.

    • The lagrange command: run, validate, graph, determinism, info.

    • Python bindings (import lagrange).

  • ROS 2.

    • /clock, /tf, /odom, /joint_states, /scan, /points, /imu, /cmd_vel.

    • The simulation_interfaces services.

    • ros2_control in lockstep (lagrange_ros2_control/LagrangeSystem).

    • The Nav2 warehouse demo.

  • Formats. URDF and xacro, SDFormat, MJCF and USD robots and worlds, with a vendored OpenUSD 26.08 and the first schemas (sensors); glTF, OBJ, STL and Collada meshes.

  • Docs. A user guide and an AMR tutorial that runs end to end (tools/ci/tutorial_amr.sh). Benchmarks with recorded baselines.

  • Packaging. .deb packages built with bloom (dry run), and a container image that runs the Nav2 demo headless.

Acceptance criteria (roadmap §3)

#

Criterion

Status

Evidence

1

Nav2 through lagrange_ros2_control, 10 simulated minutes, waypoints reached, no collision

not met yet

The patrol works once Nav2 is up: every launch that came up reached its waypoints without a collision (1-minute runs on the host and in the container; the ros2_control branch reached 54/54 waypoints over 10 minutes before the rebase). On 2026-10-25 the 10-minute run with ros2_control did not get past Nav2’s start-up in two attempts (known issue below)

2

The v0.1 viewer connects with no protocol major change and shows scene, signals, domains

met

Protocol v1 unchanged; screenshots in docs/guide/tutorial-amr.md

3

A unit mismatch fails at startup naming both modules, the signal and both units

met

test_scheduler (UnitMismatchNamesBothModules...), tutorial step 5

4

Disabling any domain with its substitute keeps the simulation running

met

test_toggles (#48)

5

Deterministic mode: two runs bit-identical for 10 s

met

lagrange determinism (601 identical hashes over 60 s); test_determinism_check

6

Lidar Tier 0 and Tier 1 agree within sensors.md §7

met

test_lidar_tiers (worst 0.16 mm on Iris Xe, 0.31 mm on llvmpipe; same shape 100 %)

7

URDF, SDFormat and MJCF versions import to USD and simulate

met

test_usd_assets, TheSdfAndMjcfVersionsDriveLikeTheUrdf

8

A world 500 km out simulates identically to one at the origin

met

FloatingOrigin.AWorld500KmOut... (#51)

9

Real-time factor ≥ 1.0 for the demo with the lidar on Tier 1

met

39× on Iris Xe while driving, headless (v0.2.0-comparison.md); not yet measured on an RTX desktop

Moved to v0.3, with reasons (docs/plan/progress.md)

  • G2 overlay files: they need the electrical schemas.

  • B8 collision bake: it needs mesh colliders and CoACD.

  • E6 scene editing: its test edits a motor parameter.

  • E8 replay in the C++ server: comes with the replay substitute.

  • H7 thin electrical slice: v0.3 builds the full electrical domain.

  • C5 runtime rebase: deferred to v0.7 by the owner.

Known issues

  • Nav2 start-up: about half the launches on the development machine have one ROS process never discover another’s endpoint. Once up, every patrol reached its waypoints without a collision. Mitigated by retries and one relaunch in nav2_demo.sh; root cause not found (docs/plan/progress.md).

  • Not run while GitHub Actions are disabled:

    • the ASan/UBSan build of the USD packages (stopped by the system for low memory);

    • the nightly jobs (benchmarks, determinism, tutorial, Nav2).

    Everything else was verified locally.

Comparison with Gazebo Jetty

See v0.2.0-comparison.md: the same AMR and warehouse, with method and limits.

Release checklist

  • [x] Acceptance criteria 2 to 9 met, with evidence above

  • [ ] Criterion 1: a 10-minute Nav2 patrol through ros2_control, which needs the start-up issue fixed or a machine where it does not occur

  • [x] Release notes and comparison drafted

  • [x] Gated changes merged under the autonomous mandate listed for review (progress.md)

  • [ ] Owner review of the gated merges (#46 and the others in the mandate table)

  • [ ] GitHub Actions re-enabled and green on dev (gh workflow enable ci and the others)

  • [ ] ASan/UBSan of the USD packages

  • [x] Released: release/0.2 merged into main and tagged v0.2.0 on 2026-10-25 at the owner’s request (release-please could not run with GitHub Actions disabled; its PR #33 was closed)