Repository navigation
OmniSim v9.2.1
More of the same work as v9.2.0: robot makers' own published models, run as they ship, kept
exposing places where OmniSim read them differently from their own files or from plain MuJoCo.
Each fix below was measured against plain MuJoCo or against the documented rule, has a pin test
that fails on the old engine, and (except where noted) a value-parsed hatch that restores the old
behaviour. The engine and the physics runtime both change, so use the v9.2.1 installer (doctor
must read bundle == source).
⚠️ Behaviour changes — read before upgrading
- URDF links with
<inertial>but no<collision>no longer collide. The importer gives such a
link a 1 cm placeholder sphere so it has a body; that sphere used to collide, so every
collision-less link carried an invisible marble. URDF semantics, and MuJoCo's own URDF import,
give it no contact. Measured: a LimX HU_D04 arm stopped against a box at −44° where MuJoCo's
reaches −67.6° (now −67.57°); TRON1 stood on the marbles under its feet with 12.03 N·m at the
knee (MuJoCo 0.54; now 0.5425519). The DJI Mavic now rests on its own body box. Hatch:
OMNISIM_NEWTON_URDF_PLACEHOLDER_COLLIDES=1(this one is opt-in to restore). - Mesh colliders get full convex hulls and multi-point contact by default. Hulls were capped at
64 vertices and each mesh pair produced one contact point, so grasps on meshes slipped where
MuJoCo's held (the Elephant F100 lift now holds, 8.47 cm, the same as MuJoCo). Across the 56
catalogued worlds with mesh colliders wall time went 709 → 706 s, two verdicts went FAIL → PASS
and none got worse. Hatches:OMNISIM_NEWTON_MULTICCD=0,OMNISIM_NEWTON_MESH_MAXHULLVERT=0
(or=nfor an n-vertex cap). - URDF
<mimic>is imported. It was dropped, so mimic-driven grippers (Robotiq 2F-85/2F-140,
Franka Hand, Kinova, LimX, Dexmate, Elephant) arrived as independent motors and the fingers
drifted apart under load. The follower's motor now names its leader (RotationalMotor/
LinearMotormimicMotor,mimicMultiplier,mimicOffset) and the engine enforces it as a
MuJoCo joint-equality constraint; commands sent to a follower are ignored. Elephant F100 with its
five mimics: lift 8.70 cm at a 0.714 rad drive, plain MuJoCo with the same equalities 8.82 cm at
0.725. Hatch:OMNISIM_URDF_MIMIC=0. - A
<rest>pose on a joint whose range passes ±π is read correctly. Since v9.0.0 such a joint
spawned posed at<rest>while its sensor read 0 there, so every command landedrestradians
off (a UFACTORY xArm6 joint3 was driven to −245.6°, past its −225° stop). This changes the
shipped quadrupeds whose thigh range passes π: the Unitree Go2 and B2 and the OmniQuad (28
joints) no longer crouch at twice their stance angle — holding its standing pose the Go2 now
stands level at 0.310 m instead of slumping to 0.184 m. The Go2/B2 RL deploy worlds still fall:
they run a 16 ms step without the quadruped stance recipe (substeps 8, compound colliders,
ground friction 2), on the old engine and the new. No hatch (the old state was inconsistent). - A body with Physics under a Robot without Physics is fixed to the world, as
docs/reference/physics.md(inherited from Webots) has always said. The physics-less Robot used
to become a dynamic body that rolled the child's mass up, and the physics-less "larger static
base" form failed to load when its joints hung off a child Solid. Measured on a published Webots
twin (pib): its body dropped 22.6 mm onto its desk, slid 0.40 m and fell off; now it holds its
spawn pose and the replay is identical to the hand-pinned model. Of 223 catalogued worlds one
changes:warehouse_industrial's ConveyorBelt no longer sinks 0.68 m through the floor. Hatch:
OMNISIM_NEWTON_PHYSICSLESS_ANCESTOR_WELD=0.
Fixed
- A moving joint on an empty fixed frame loads.
body -fixed-> mount (no <inertial>) -continuous-> wheel(Husarion Lynx) ortorso -fixed-> arm_center -revolute-> arm(Dexmate
Vega) FATALed with "parent body ... never registered a Newton body". The joint now attaches to
the body that carries the frame, bitwise identical to the same robot with the frame folded in.
Hatch:OMNISIM_NEWTON_JOINT_PARENT_FRAME_WALK=0. - An empty link on a moving joint carries the links fixed below it. A link with neither
<inertial>nor<collision>(Kinova Gen3 lite'send_effector_link, gripper underneath) had no
body, so the fixed subtree registered as a separate free body and fell away; the gripper's mass
vanished from the arm's joint torques. Gen3 lite joint_2 now 13.719 N·m with a 0.5 kg payload
(plain MuJoCo 13.712). Hatch:OMNISIM_URDF_EMPTY_LINK_BODY=0. - A prismatic joint's motor reports its travel.
LinearMotor.getMinPosition()/
getMaxPosition()read 0/0 while the stops enforced the URDF range (Niryo Ned2 jaws: now
−0.01/0.01). Solver limits are unchanged; asetPosition()past the travel is now clamped with
a warning, as for revolute joints. No hatch. - No false "collisions will have no effect" warning for fixed-child colliders. Since v9.2.0
those colliders are attached and do collide; the inherited Webots warning still fired and was
read as "this robot has no caster" (Ubiquity Magni). It now fires only with
OMNISIM_NEWTON_FIXED_CHILD_COLLIDERS=0, when it is true.
Full Changelog: v9.2.0...v9.2.1