Skip to content

OmniSim v9.1.0

Choose a tag to compare

@omnilink-tech omnilink-tech released this 30 Sep 03:38
· 25 commits to main since this release

Physics corrections that change what some worlds do, a harness that reports the
joint limits the solver actually enforces, OmniLink bridges that can hold a
30-minute operations shift, a PX4 x500 drone, and a legged OmniLink demo on real
contact physics. Four of the physics and harness fixes came from external users'
reports.

⚠️ Behaviour changes — read before upgrading

  • A staticBase robot's root now collides. A URDFRobot with
    staticBase TRUE (or an articulated Robot whose root has no Physics) was
    welded to the world with no collision shape: grippers, props and other
    robots passed silently through every fixed robot base, and MuJoCo generated no
    contacts for those pairs. Found through an external 5-DOF arm evaluation, where
    the gripper sat 17.7 mm inside its own base. The root's boundingObject and the
    colliders of its joint-free fixed children now attach to the welded body.
    Checked on all 78 tracked staticBase worlds (load + --fail-on-runaway):
    verdicts are unchanged apart from two timing flips on one world pair that both
    pass at 40 s. Physics is bit-identical in 37 of the 38 worlds with tracked
    dynamic bodies, and the omniarm6_real_pick_place grasp is bit-identical.
    OMNISIM_NEWTON_STATIC_BASE_COLLIDERS=0 reverts (value-parsed). Pinned by
    tests/test_newton_static_base_colliders.py.
  • A body whose mass comes from density now weighs what its density says.
    Every dynamic Solid defined by density (Physics { density 250 }, or a bare
    Physics {} at the default 1000 kg/m³) was registered on the Newton solver at
    0.25 kg, whatever its size: the body build summed the raw mass field
    (−1 for these bodies) and fell through to a 0.25 kg fallback. Measured: a
    0.2 m box at density 250 → 2 kg, a 0.2 m-radius sphere at density 2000 →
    67 kg, a 0.5 m box with Physics {} → 125 kg, all previously 0.25 kg. Bodies
    that declare mass (every URDF link and every shipped demo) are unchanged.
    Most visible for objects spawned at runtime and then simulated with
    /sim/rebuild_physics: a heavy ball used to bounce off a stack of light
    blocks. Affected shipped worlds are samples only (the device samples,
    physics_primitives, cylinder_stack); all still load, and the three with
    top-level bodies pass --fail-on-runaway. Pinned by
    tests/test_newton_density_mass.py.
  • GET /robot/<def>/joints reports the limits the physics enforces. It read
    lower/upper from the joint's raw stops only, but the solver uses the
    Motor's minPosition/maxPosition when they are set. So every URDF revolute
    limited by its motor alone read lower: 0, upper: 0, and neither hit_limit
    nor the joint.limit_hit event could ever fire for it. Reported by an external
    user on a Galbot R1. lower/upper are now the effective limits, with a new
    limit_source ("motor", "stops" or null) and the raw
    stop_lower/stop_upper and motor_lower/motor_upper beside them.
    An unconstrained joint now reads null, not 0/0. joint.limit_hit
    uses the same limits and carries limit_source. PROTOCOL.md §7.14 and §10.4
    and the MCP get_robot_joints description are updated.

Fixed

  • Two joint-registration defects on Newton, both reported by an external
    user with minimal probes. A joint under a fixed joint that merged a rotated
    parent into its leader body turned about the wrong axis (0.353 rad
    orientation error at q = 0.25 rad; 3.7e-6 rad after). A slider child authored
    at an angle was snapped to its parent's orientation. Across the 72 shipped
    robot URDFs (743 moving joints) no axis changes; the M20 Piper's two gripper
    fingers now keep their authored orientation.
  • The first demo no longer fails on a machine with no CUDA GPU. A world
    pinned to newtonSolver "mujoco_warp" JIT-compiled its kernels for the CPU,
    and the first step landed after run-headless --until-finalized's 10 s
    budget, so the quickstart's warehouse_husky FAILed on a healthy install.
    With zero CUDA devices the pin now runs the CPU MuJoCo step and the solver
    label says it fell back (OMNISIM_NEWTON_MJWARP=1 forces the old path).
    Reported by an external tester.
  • run-headless printed the --step-wait-timeout budget as the time it waited;
    it now reports the measured wait beside the budget.
  • The Windows build's Python probe embedded the newest CPython it found (3.14
    beside 3.12 won silently). It now prefers Python 3.12, warns on any other
    pick, and make linker-info prints the choice.
  • Harness mutation verbs emit events. POST /robot/<def>/joints/set and its
    siblings now poll the contact, grip, joint-limit and damage trackers during
    their settle steps, as /sim/step does; motion inside them used to fire no
    contact.* or joint.limit_hit events.
  • The Mavic bridge published v_xy/v_z over wall-clock time, so below real
    time (e.g. under 1080p capture) a moving drone read ~0 m/s, a flight tool
    "settled" mid-flight and landings overshot by up to 0.47 m. Velocity is now
    taken over simulation time (landing drift 0.35–0.47 m → 0.018 m on the x500).

OmniLink bridges: a robot that keeps a 30-minute shift

Built from what a new benchmark of robot operations under change (below) found
the bridges getting wrong. Everything here applies to the mobile bridge; the
gate still vets every motion, including scheduled ones at the moment they fire.

  • Stop means stop, now. A "Stop!" mid-plan used to queue behind the running
    motion (answered after 9.4 s, after the robot finished its leg and turn); it
    is now answered in ~0.3 s. "Emergency stop!", "Abort!", "Stop turning!" parse
    as stops; "Stop at the door." does not.
  • A new order while the robot works stops the work first, and the reply
    says so. Remarks, information and questions are answered beside the running
    work instead of interrupting it.
  • Spatial rules and keep-out zones. "Keep your x at or below 0.40 until I
    say so" and rectangular keep-out zones are recorded, enforced on every drive
    (a drive stops short of the line and reports that it was clipped) and liftable
    only by the supervisor, the safety lead or whoever set them; a rule worded as
    permanent cannot be lifted.
  • Route planning. drive_to plans around zones, lines and obstacles the
    robot has met; a blocked leg is detected from speed within a fraction of a
    second, backs off and replans, and a stalled drive ends as blocked instead
    of pushing until its timeout.
  • Orders for later. "In 15 seconds, drive forward 1 metre", "if anything
    bumps into you, back away 0.3 metres", and timed orders to named places run on
    the simulation clock. They are not persisted: nothing starts moving by itself
    after a restart.
  • Replies come from what was measured. A parsed order is answered after the
    motion, from the robot's measured result ("I stopped short: I drove +1.20 m of
    the +2.00 m asked"), not by echoing the request.
  • Shift memory. Named places, visit counts, people and roles, rules and the
    record of scheduled orders are kept per shift; /state and the new
    get_shift_memory tool expose them to every caller. A radio-style
    "Name: ..." prefix is read as the speaker.

Drones and URDF import

  • PX4 x500 (projects/robots/px4/x500/, BSD-3-Clause, generated
    deterministically from the PX4 gazebo models; full trail in its
    PROVENANCE.md) and showcase/x500_arena.omniworld, flown by the Mavic
    OmniLink bridge. Needs OMNISIM_URDF_USE_SENSORS=1, like the Mavic chat demo.
  • Opt-in flight model per airframe: rotor spin-up lag, ground effect, drag
    and seeded, modelled wind (not measured air). /capabilities.airframe states
    which effects a world declares; the Mavic declares none and is unchanged
    except for a softer landing flare.
  • The URDF importer reads <texture>, colour alpha and material names
    (textures, transparency and appearance names were parsed and dropped).

Demos and capture

  • X30 plant inspection (showcase/x30_plant_inspection.omniworld, bridge on
    port 8796): a Deep Robotics X30 on an inspection round, driven through
    OmniLink, trotting on real contact physics via the quadruped bridge's new
    opt-in --locomotion crawl|trot. It is a scripted gait on flat ground with
    Supervisor pose and no obstacle sensing, and the world says so. Everything
    behind the demo lives in agents/omnilink_demos/.
  • Capture records on simulation time: POST /capture/record/{start,stop,status}
    saves a frame every period_ms of sim time, so footage plays back smoothly
    at any real-time factor, and record/follow keeps a camera on a robot.

Benchmarks

  • python -m omnisim ops-bench (validate, run, judge, regrade)
    grades what happens while a robot works: standing rules, interruptions,
    timed orders, disturbances. See tests/benchmarks/robot_ops/.
  • Warehouse shift v1 results (README, "How OmniLink compares in OmniSim"):
    one blind 30-minute shift, seven configurations, three repeats each, one
    laptop, simulation only. The report states its deviations first:
    tests/benchmarks/robot_ops/evidence/SHIFT_V1_RESULTS.md, with costs in
    SHIFT_V1_COSTS.md. The raw per-episode evidence is published as a
    downloadable bundle on the benchmark page rather than in this repository.

Known limitations

  • The OmniLink edge connector still needs OmniLink SDK 0.6.4, which is not yet
    on PyPI; with the published 0.6.3 it reports itself unavailable.
  • Bump detection runs only while the robot is at rest, and its 5 mm threshold
    was chosen on one fixture.