Repository navigation
OmniSim v9.2.0
URDF robots now arrive in the solver the way their files describe them, and a joint resting on
its limit stops hiding load from the joints above it. These defects all showed up the same way:
while we ran companies' and labs' own published robots, the importer, not their models, kept
breaking the runs. Every change has a value-parsed hatch (=0 restores the old behaviour, except
where noted) and a pin test. This release changes the engine, so use the v9.2.0 installer; the
physics runtime changes too (doctor must read bundle == source).
⚠️ Behaviour changes — read before upgrading
- Fixed-joint offsets are kept on import. When the importer skipped an empty root link such as
base_footprint, it dropped that fixed joint's offset, so every child joint moved by it (the
TurtleBot3 Burger's wheels sat at z 0.023 instead of 0.033 and the chassis rode ~1 cm high).
Hatch:OMNISIM_URDF_REROOT_KEEP_OFFSET=0. The ROSbot and ROSbot XL chat worlds spawned high to
compensate; their spawn heights are corrected (0.05 → 0.0075, 0.08 → 0.005). - Merged bodies carry the real centre of mass and inertia. Links joined by fixed joints were
merged with their summed mass placed at the leader's origin and only the leader's inertia. The
composite calculation is now on by default (OMNISIM_NEWTON_COMPOSITE_INERTIA=0reverts; that
variable used to switch on when merely present, it is now value-parsed). Turn ceilings of the
chat bases moved with it and the bridge table is re-measured: Husky 0.520 → 0.475, TurtleBot3
Burger 0.942 → 0.912, Waffle / Waffle Pi 0.955 → 0.933 (Jackal, ROSbot, ROSbot XL unchanged). - Colliders on fixed child links are registered. Casters, feet, gripper palms and sensor
housings attached by fixed joints never collided; a robot rested on whatever its first collider
was (the TurtleBot3 rocked about its axle indefinitely). They now attach to the merged body at
their relative pose, independent ofnewtonRobotColliders. Hatch:
OMNISIM_NEWTON_FIXED_CHILD_COLLIDERS=0. The OmniArm6 2F-85 pick-and-place was tuned around the
missing palm (its descent pressed the palm 20 mm into the block); its grasp height is now 0.275
(PICK_GRASP_Zoverrides) and it passes its strict carried + pinched + placed check, which it
did not before. - A joint resting on its limit transmits load. A post-step clamp snapped joints back to exactly
the limit, where MuJoCo applies no limit force, so the links beyond a stop fell freely each step
and their weight never reached the parent joint (OpenArm elbow on its stop: the shoulder read
3.84 N·m against a true 11.68). A slack band (0.05 rad / 0.005 m) is now left to MuJoCo's soft
limit, and sensors still read the clamped value. Hatches:OMNISIM_NEWTON_JOINT_CLAMP_SLACK=0,
OMNISIM_NEWTON_JOINT_CLAMP_SLACK_LINEAR=0. - Mobile-base import fixes. Revolute wheels declared with huge limits (±1e16) are treated as
continuous and turn under a velocity command (OMNISIM_NEWTON_UNBOUNDED_LIMIT_AS_CONTINUOUS=0);
a collision mesh's<mesh scale>is honoured through a newMesh.scalefield
(OMNISIM_URDF_COLLISION_MESH_SCALE=0); the 1 mm placeholder sphere no longer collides once the
robot has real colliders, so robots whose origin sits below their wheels stop resting on it
(OMNISIM_NEWTON_WRAPPER_PLACEHOLDER_COLLIDES=1restores it);webots://URLs in upstream PROTOs
resolve likeomnisim://(OMNISIM_LEGACY_WEBOTS_SCHEME=0); andsetAvailableTorque(0)now
frees a joint (OMNISIM_NEWTON_AVAILABLE_TORQUE=0). Robotnik's own Webots demo world goes from 16
load errors to none. - Joint damping and friction are wired through, but OFF by default.
dampingConstant/
staticFriction(and URDF<dynamics>) now reach the solver with
OMNISIM_NEWTON_JOINT_DYNAMICS=1. They ship off because a velocity servo's stall torque is still
capped by its gain clamp rather than the declared effort, so a declared wheel friction can stall
a robot whose real motor would drive it (a Raspberry Pi Mouse fell to 0.16 of commanded travel),
and worlds calibrated without damping drift (the metazoa reef roamed at half speed).
Known issues (not caused by this release, still open)
- The shipped walking demos fall on their documented deploy paths, in v9.1.3 as in v9.2.0.
Go2 and OmniQuad: since the 2026-09-20 change that stopped clamping URDF joint ranges at ±π, a
thigh range that crosses +π starts the leg in the wrong pose. G1: since plain static colliders
moved onto Newton's world body (2026-09-07), every body index shifted and the deploy hook's
hard-coded harness body now lands on a leg. Clamped URDF copies and a by-name body lookup restore
them in testing; the fixes come in a follow-up release. - The Go2 rough-terrain baseline policy trips on its new feet. With fixed-child colliders it
stands on its foot spheres for the first time; trained on the footless model, it covers 6.1 m
before falling instead of 8.1 m.OMNISIM_NEWTON_FIXED_CHILD_COLLIDERS=0restores the trained
model until it is retrained.
Docs
- The README's limits section and
SPONSORS.mddescribed pause and breakpoints as missing; both
shipped in v9. They now describe the held pause and break-on-event with their two latency regimes,
and keep the honest list of what is still missing (watch conditions, record, replay, run-diff, a
real checkpoint). Light mode silences 5 of 11 event types, not 5 of 10.
Tests
New: tests/test_newton_urdf_fixed_links.py (engine), tests/test_newton_joint_limit_load.py,
tests/test_newton_joint_passive_dynamics.py, tests/test_newton_mobile_base_fixes.py. Verified on
this release candidate: the unit lane, the smoke worlds, the lidar and static-base engine pins, and
old-versus-new runs (same binary, hatches off) across the wheeled chat demos, the Deep Robotics
showroom, the X30 terrain crawl, the M20 + Piper arm and the OmniArm6 pick-and-place.
Full Changelog: v9.1.3...v9.2.0