Skip to content

OmniSim v9.2.2

Choose a tag to compare

@github-actions github-actions released this 09 Oct 18:46
· 17 commits to main since this release

Fixes found running robot makers' published models and a third-party benchmark on the installed
v9.2.1, plus the first outside fix to the ROS 2 bridge. Each fix has a pin test that fails on the
old engine; the two behaviour changes carry a value-parsed hatch that restores the old behaviour.
The engine and the physics runtime both change, so use the v9.2.2 installer (doctor must read
bundle == source).

⚠️ Behaviour changes — read before upgrading

  • Every collider of a multi-shape boundingObject registers by default. A Group of offset
    primitives, or a URDF link with several <collision> elements, collided as its first shape only
    unless WorldInfo.newtonCompoundColliders was TRUE, and that field also selects the inertia
    source, so its default could not change. Leo Rover's wheels rolled on the 0.057 m hub listed
    before the 0.0625 m tyre; a Chair collided as a seat with no legs. Collider registration is now
    separate from inertia: a body with more than one shape registers all of them, a single-shape body
    is unchanged, and newtonCompoundColliders keeps only its inertia tie-break. A new
    WorldInfo.newtonCompoundShapes field (default TRUE) pins one world to the old behaviour. The
    20 shipped B2 and Go2 policy worlds set it to FALSE: their policies were trained with only the
    foot sphere of each calf colliding, and with the shins colliding too the B2 walk deploy covered
    0.48 m instead of 2.37 m in 8 s. Measured old against new: the pinned B2/Go2 worlds and the X30,
    Lite3, G1 and 8-Husky worlds are bit-identical; a bin-picking part and the warehouse forklift end
    148 mm and 59 mm away. Hatch: OMNISIM_NEWTON_COMPOUND_SHAPES=0.
  • A Cylinder collider off the joint graph is a real cylinder. Every Cylinder was a capsule
    stand-in, so a disc, puck, table top or chassis plate (radius at least half the height) rested one
    radius above where it was authored: a r=0.08 h=0.01 disc at z=0.0846 instead of 0.005. Props,
    static geometry and robot roots now collide as true cylinders and rest at the authored height to
    under 0.2 mm. Wheels and arm links keep the capsule, because there it is not neutral: an open-loop
    Husky pivot turned 132° on capsules and 82° on true cylinders. The Husky pivot and the 8-Husky
    drive are bit-identical; a sweep of the 61 hand-authored wheeled worlds keeps the same 22 rolling.
    Hatches: OMNISIM_NEWTON_CYLINDER_NATIVE=0 (capsules everywhere) or =1 (true cylinders
    everywhere, wheels included).

Fixed

  • An omnisim:// asset missing from an installed release is fetched from that release on
    GitHub.
    The installer ships no PROTO files, and a world written with omnisim:// PROTOs, the
    form the docs recommend, failed to load on an install: the fallback pointed back at the missing
    local file (since 2026-04-11). Found running a ghostloop backend on the v9.2.1 installer.
  • A Collada file that instances geometry loads instead of crashing. The mesh loader sized its
    arrays per mesh but wrote each mesh once per node that used it, so a file reusing a mesh overran
    them and the engine died with no log line. Leo Rover's Chassis.dae (27 geometries in 35
    instances) crashed every load of the robot; it now loads.
  • A mesh path longer than 260 characters loads on Windows. Mesh, CAD-shape and cloth files
    open through the Windows long-path form; a URDF package unpacked a few folders down used to lose
    that geometry silently.
  • ROS 2 Tier-2 /cmd_vel streaming no longer blocks. The client called the mobile bridge's
    /set_velocity without wait, whose default waits for an achieved-rate measurement, so each
    command blocked the single-threaded executor and tripped the watchdog: a Husky Nav2 goal took
    197 s of wall time with 6 recoveries, now 10 s and none. Thanks to @LingZhen07 (#22, #23).
  • The joint-damping log line says whether damping is applied. URDF <dynamics damping friction> apply only with OMNISIM_NEWTON_JOINT_DYNAMICS=1 (default off), but the log reported
    every declared value as applied. It now says "applied" or "NOT applied".

New Contributors

Full Changelog: v9.2.1...v9.2.2