Repository navigation
OmniSim v9.2.4
Three physics-runtime and importer fixes found running research labs' and robot makers' published
models, and an outside AMCL localization benchmark. Each fix has a pin test that fails on the old
engine and a value-parsed hatch that restores the old behaviour. The engine and the physics runtime
both change, so use the v9.2.4 installer (doctor must read bundle == source).
⚠️ Behaviour changes — read before upgrading
- A URDF
<inertial>origin'srpyrotates the inertia tensor. Both URDF importers (the
engine's andscripts/dev/urdf_import.py) used only the origin'sxyzas the centre of mass and
dropped itsrpy, so a tensor written in a rotated inertial frame reached the solver unrotated:
mass and centre of mass were right, the rotational inertia about each axis was not. The tensor is
now expressed in the link's axes (I' = R I R^T, URDF fixed-axis roll-pitch-yaw), as the URDF spec,
Gazebo and PyBullet do. Plain MuJoCo 3.11's own URDF import drops therpyas well, so on these
models OmniSim and plain MuJoCo now differ. Probe: a 1 kg pendulum whose inertial frame is yawed
90°, released horizontal, fell 8.7 mm in its first 100 ms, like a body with the tensor unrotated;
it now falls 45.3 mm, exactly like the same body written pre-rotated (PyBullet: 43.7 mm for both).
Shipped robots that change: UR3e, UR5e and UR10e (3 of 7 links; tensor entries move by up to 39% of
the link's largest moment), the Trossen ALOHA arms (9 of 10 links, up to 89%), and the Deep Robotics X30
(three hips rolled by π, which only flips the signs of two products of inertia). Hatch:
OMNISIM_URDF_INERTIAL_RPY=0. - The joint speed cap no longer pushes a free-floating robot around. After every step the
physics runtime caps each joint at itsmaxVelocity(URDFvelocity), and zeroes a joint's
velocity into a stop it has overrun, by overwriting that joint's velocity. On a fixed base the
world takes the reaction; on a floating base nothing did, so the momentum the cap removed came
back as base motion. A free-falling UR10e whose saturated shoulder-pan servo was capped spun at
10.7 rad/s about z within 25 ms and slid 2.1 m sideways in 0.5 s, where plain MuJoCo just falls.
The robot's floating base now takes the reaction: the correction holds the robot's linear and
angular momentum and leaves every other joint's velocity as before. The same UR10e now falls
straight (base 0.0 mm sideways at 0.5 s, wrist below 0.13 rad/s), and a two-link probe's angular
momentum ends at 5e-4 kg·m²/s instead of 3.3. Fixed-base robots are bit-identical. On
mujoco_warpa floating robot's joints are not speed-capped instead. Hatch:
OMNISIM_NEWTON_CLAMP_MOMENTUM=0. - A motor with zero torque is not held to its speed limit. v9.2.3 made a motor with zero
available torque passive, but the runtime's post-step cap still held its joint tomaxVelocity,
which models a motor's top speed; an off motor has none. Found re-running Dexmate's Vega-1u: its
passivetorso_flip(URDF effort 0, velocity 0.1, 15.7 kg above it) crept onto its stop at
0.11 rad/s for 9.5 s, where plain MuJoCo drops it in 0.4 s. A joint whose available torque is 0
is now exempt from the speed cap (its range limits still apply); restoring torque restores the
cap. Hatch:OMNISIM_NEWTON_PASSIVE_VEL_CLAMP=1keeps the cap.
Added
- A manual Husky AMCL localization benchmark (
tests/benchmarks/husky_amcl_localization/):
stock ROS 2 Jazzy AMCL on the known-wall Husky world, with initial and post-motion stationary
localization errors measured against independent harness poses, convergence and jitter, and a
sequential cold-start repeat runner. AMCL gets simulator ground-truth odometry, so it does not
measure robustness to wheel-odometry drift. It needs Linux or WSL2 with ROS 2 Jazzy. Thanks to
@LingZhen07 (#28).
Known issues
- Two tests fail, and neither is caused by this release's fixes:
tests/test_newton_ik_slots.py::test_reachable_target_is_actually_solved(a floating-base IK
target is solved to 0.157 m instead of 1 mm; identical with the v9.2.3 runtime) and the engine
testtests/test_urdf_empty_link_carries_fixed_subtree.py(2.4 mm against a 1 mm tolerance;
identical with both new hatches set).
What's Changed
- tests: add measured Husky AMCL localization regression benchmark by @LingZhen07 in #28
Full Changelog: v9.2.3...v9.2.4