Lagrange v1.4.0: build a robot in the viewer (release notes)

v1.4.0 is a feature release on the stable 1.x line. The schemas, plugin ABI, protocol, profiles and recordings changed only by addition (RFC 0007); the behaviour changes are listed under “Upgrading from 1.3.x” below.

Status: released on 2026-11-15 as v1.4.0, on the owner’s request. Documentation: https://open-physical-systems.github.io/Lagrange/.

Install

curl -fsSL https://raw.githubusercontent.com/Open-Physical-Systems/Lagrange/main/packaging/install.sh | bash
source /opt/ros/lyrical/setup.bash
lagrange amr

The robot builder (RFC 0016)

Press Ctrl B in the viewer (or the Launcher’s Build tile) to make or edit a robot without writing a file by hand. Guide: docs/guide/builder.md.

  • Start from a template, a URDF or a xacro file. Add links, joints and shapes; move and turn them with gizmos; mass and inertia come from the shape and a material’s density.

  • Meshes as parts (STL, OBJ, DAE, glTF), with a collision choice per link: the mesh’s convex hull, convex parts split by CoACD, template shapes fitted to the mesh, or shapes you draw.

  • Sensors from the catalog, placed on the part, with each parameter’s range checked.

  • Power: motors, battery, bus, converters and fuses, written as an overlay.

  • Live validation (Problems), undo and redo, autosave, and a prompt before unsaved work is lost.

  • Test runs the robot on a server of its own, in an empty world, a sample world or a world file, at a start pose you pick; joint sliders drive it; Back returns to editing.

  • Saves a URDF and its overlay; exports USD.

Viewer

  • Every workspace has its designed tabs: Domains (modules, coupling, diff, schema), Electrical, Environment, Signals, Replay (compare, sim vs real, validation), ROS 2, Performance (KPI cards, threads, GPU passes, memory), World and Scene; a new Launcher; the command palette (Ctrl K); the 4D radar point cloud drawn.

  • Text at the design’s size (1.3× larger than 1.3.0), cut with “…” and a tooltip where it does not fit, no overlapping labels, down to 1024 × 640.

Physics and worlds

  • Jolt: sub-steps with joint reactions and contact forces averaged over the sub-steps; friction per collider.

  • MuJoCo: noslip_iterations profile key, joint armature from MJCF, static scenery as world geometry (faster).

  • One ground per run: Lagrange’s plane only when no file brings a floor.

  • Tethers: the rotors’ downwash blows the cable (tether/<n>/downwash); RFC 0011 complete.

  • OpenDRIVE: superelevation (banked roads), lane heights (kerbs, sidewalks), level lanes, and raised roads vehicles drive on; sample world assets/worlds/banked_ring.

  • Cars: tyre forces in the road plane with the normal load along the road normal; the brake holds a car on a slope; traffic vehicles keep clear of the robot and wait at a junction it is crossing.

Python and learning

  • lagrange.envs.ParallelEnvs over shared memory (no pickling per step); environments take profile=, overlay= and world=.

  • command_joint returns a JointCommand saying whether the joint took the command, why not, and what was clamped.

Quality

  • Four QA rounds, two code reviews and an ASan/UBSan/TSan pass; every finding fixed or written down. Replays of robots with cameras run; /spawn_entity follows allow_renaming; states saved by 1.1 to 1.3 restore again.

  • Tests and CI scripts run side by side (a private temporary folder per test, ROS tests on the caller’s ROS_DOMAIN_ID).

Upgrading from 1.3.x

The full list is in CHANGELOG.md under 1.4.0, “Breaking or behaviour changes”. The ones most likely to matter:

  • Python command_joint returns a lagrange.JointCommand instead of a bool: if, not, == True and == False behave as before; is True and JSON serialisation do not. Targets beyond the robot file’s limits are clamped (Python and the viewer; ROS /joint_commands is not).

  • /spawn_entity renames on a clash only with allow_renaming: true; otherwise a taken name fails with NAME_NOT_UNIQUE, text naming no robot with NAME_INVALID.

  • Kind parameters outside the catalog’s range refuse the run instead of being clamped.

  • One ground per run: a world or robot file with its own floor no longer gets Lagrange’s plane on top of it.

  • State hashes change for MuJoCo runs with static scenery, Jolt runs, cars (tyre contact frame, brake) and traffic; runs stay deterministic.

  • Viewer text is 1.3× larger; headless screenshot scripts with fixed click coordinates may need new ones.

  • The quadruped example’s robot moved to assets/robots/quadruped/quadruped.urdf.

Checked

Whole workspace (66 packages) built and tested on Ubuntu 26.04 / ROS 2 Lyrical at dev #423: 2506 tests, 0 failures after #424 (a builder test that clicked a sample world by position when banked_ring joined the list); lagrange determinism identical for the AMR, sedan, quadcopter and quadruped profiles; every example and tutorial passed at #408, and the sanitizers (ASan, UBSan, TSan) found nothing left in Lagrange code.