Known limits

What does not work yet, one line each. Every line links to a page with the measured size of the limit and what to do instead today. Read the lines for your area before you trust a result there.

Rigid bodies and contact

• A fast small body passes through a thin static part: a 2 cm steel ball lands on a 1 cm plate from 8.9 m/s and passes through from 9.9 m/s.
• A box sliding at metres per second hops and rocks: a cube at about 2.5 m/s rises a few millimetres and pitches 1-2 degrees (up to 7 for a small cube).
• A body on collision="sdf" or "mesh" never fully rests, and fails under about 40 g or from 40 kg.
• Light slender parts such as pens and pencils spin and roll slower about their thin axis, because their inertia is raised at build.
• With more than 8 free bodies on a GPU, a body spinning in free flight turns only to first order; the CPU's default route keeps rk4 at 3-6x the cost per step.
• Stacks of free bodies step 2.4x slower by default than stock Newton MuJoCo, and the opt-in fast mode that closes the gap softens contacts.
• The opt-in Newton 1.8 rigid solver lets a robot arm's contacts through 54 times more often than the copy before it.
• A lightly pressed box sliding at a 1.25 ms step can tip in scenes that set their own MuJoCo contact gains; the default path slides flat.
• Mesh contacts can take the wrong inside or outside side at rare degenerate points, such as one point in a thin cup's cavity.
• A hydroelastic precision fit pushed sideways leans about 1.3 times as far as its clearance allows, with about 0.1 mm of wall penetration.
• A precision fit pushed sideways leans 2 to 6 % past its clearance, by an amount that depends on the push direction.
• A peg that misses its bore and rolls across the block slowly gains speed instead of losing it, and health stays OK.
• Rigid bodies in the cloth and soft-body solver turn by a first-order step, so an applied torque arrives short and can wrap, and a spin slows down.

Robots and grasping

• No call attaches or welds an object to a gripper: friction is the only grasp.
• Robot scenes step physics at 1.25 ms (8 steps per 10 ms period) unless you ask for a coarser step, even when a larger one would be stable.
• A robot file that is a long chain of light bodies (from 6 nested bodies) can go non-finite at its first contact on MuJoCo Warp.
• Shipped robots left with no targets drift from their start angles: Franka arms start outside a joint range and servos sag under gravity.
• The shipped Franka arms never collide link 5 with link 7 (the Panda: link 5 with the hand).
• A weld or connect in a robot file is a stiff spring, not a rigid joint, and its stiffness falls with the lighter body.
• The shipped 200-capsule bin pile starts with 54 pairs of capsules overlapping by up to 7.97 mm, so health reads DEGRADED at first.
• A body that touches cloth or soft parts is coupled through a mass that is the same in every direction, heavier than the body along its moving axis (4.5 kg for a 1.5 kg slider).

Cloth and film

• A free strip of cloth bends about 35 % too soft along a mesh axis.
• A stiff sheet (PET, nylon, polypropylene, cardboard, bending length over 5 cm) sags several times too deep between supports at the default iterations.
• A woven fabric card built from published bending data drapes differently from the same fabric on a drape tester, even when converged.
• A cloth card whose stretch stiffness is large for its mass, mesh and step stays nearly flat and creeps at about 2 mm/s, while health reads OK.
• Sheets dropped onto each other can end with their layers crossed, and health does not flag it.
• A rigid ball with a radius under about 0.7 of the sheet's cell size can go through a cloth or film sheet without touching it.
• A folded cloth keeps creeping under the default cloth accuracy instead of coming to rest within a fraction of a second.
• Folded T-shirt sleeves spring part-way back at the default mesh and contact stiffness and do not stay folded.
• Cloth held between pads by friction alone creeps out below the Coulomb limit, at about 2.9 mm/s at half the limit with default settings.
• A robot's friction pinch on cloth creeps and lets go before the Coulomb limit, so lifting, carrying or folding cloth can lose it.
• Cloth gripped by friction on a moving rigid body can slip on it right after the body reverses, moving 0.35-0.65 of the body's travel.
• Cloth resting under its own weight away from the world origin grips harder than its friction says: on a table at 0.75 m it holds or creeps slowly.
• A cloth particle that starts a step inside a rigid contact is thrown out at its depth divided by the step, 18.7 m/s for 0.5 mm at robot-finger steps.
• Cloth and soft solids with self contact move at most about 3 m/s, so a sheet dropped from 1 m falls at 3 m/s instead of 4.4 m/s.
• Self contact keeps fixed per-element buffers and can drop pairs once a scene needs more than they hold.
• Paper and film keep only a flat fold made on one mesh line: a sheet bent 90 degrees over a plate edge does not keep the fold.
• A paper or film sheet given about 3x or more of its card's thickness is launched or goes INVALID when it is folded.
• A very thin film (0.05 mm) draped over a sharp box edge goes INVALID, and the fix is opt-in in this release.
• Metal sheets sag far more than real ones at real density, which can yield them, and they bend unlike a plate across the mesh diagonals.
• A film's strain projection carries no force to bodies: a 90 degree peel of a PET strip reads 42.9 N against 8.7 N.
• The open-carton tape peel throws a light body past 20 m/s in some worlds, 1 in 40 with the task's caps and 3 of 24 without them.
• The wrapped present cannot be unwrapped yet: the task is not run and 0 % of the paper comes off.
• A two-pad pinch does not lift plies it presses hard: at 100 N the pads rise and the plies stay on the table, with no health flag.

Liquids and heat

• Liquid needs finer particles than the default in openings under 32 mm (16 mm at super), and containers under 24 mm are unreliable at both.
• A pour alone can splash a few percent over its container's rim (3 % of a 6 ml pour), and health stays OK.
• A stream poured on a fabric sheet with self contact on makes the world INVALID, and cloth takes up no water.
• Cloth, liquid and free rigid bodies lighter than twice the cloth and soft mass do not build in one scene.
• A floating container with a liquid-to-container mass ratio above 50 (such as a light cup) is refused at build.
• Liquid beside a paste or powder sees it as a solid wall: there is no wetting or percolation.
• Heated liquid does not convect: buoyancy=True (water only) gives the first push, then the flow stalls after about a millimetre.
• Boiled-off liquid particles stay in the solver and can block a heater: water at 99 degrees C on a 177 degrees C plate stops heating once 25 % of its mass has boiled off.
• A cylinder's force pressure lags when its chamber is joined to another small chamber by a wide edge, and the ringing grows.
• A syringe pushed by a constant force rebounds after its air leaves, because the plunger still carries the needle's liquid mass.
• A surface film feels only what its settings name; every other body and particle passes through it.

Cables and rods

• A cable on the chain model creeps back in a held twisting grip on the CPU, at about 0.14 % per second.
• A chain rod resting on an elastic rod sinks into it by up to the wire's radius, which can make the world INVALID.
• A thin elastic rod can fall through a soft body between its vertices when it is thinner than the gaps.
• A wire loaded close to its fully plastic moment settles slowly and less accurately.
• A rod folded tighter than one segment's length (a hairpin or coil) passes through itself and the world fails its health check.
• A thin chain rod's twist vibration rings too slowly or not at all, though a steady end torque twists it correctly.
• An elastic rod beside a robot runs far slower than the chain model on the CPU: 94.9 s against 1.00 s per simulated second.

Soft bodies, cutting and plasticity

• A body dropped on a soft solid bounces back almost as fast as it arrived, far more than the materials' restitution says.
• A small body resting on a coarse soft solid can sink into one surface particle past health's limit.
• A flat tool pressed into a soft block compresses it without limit (up to 208 % strain) because only blade edges cut.
• Cutting with copied fragments on a coarse mesh has five limits, such as a piece needing two layers of mesh nodes.
• The knife-force reference holds only until the first cut; after it the pre-split seam reads 0.65-0.95 N of contact artefact.
• Plasticity covers soft volumes, joint springs, metal sheets, wires and coil springs; bulk metal parts stay rigid.
• A soft body that yields can yield spuriously at rest when its yield strain is under about 10 %, then slump and creep.
• A punch's force on a yielding MPM solid converges only for soft solids; metals need mass scaling, else the force reads a few % of the truth.
• A drill bit that lands on an exposed bar never cuts it, however hard it is pushed.
• Drilled holes are seen only by the parts listed as pins; every other body meets the undrilled shape.
• A weld bead's start, its tack starts and a crater mid-joint are not right yet.
• Trapped air under a label is scored at 20 mm2, not at 3 mm2 inside a barcode.

Sand and pastes

• Pastes and viscous flows along the ground run slow, or leak through thin floors on the dense grid, because floors sit between grid planes.
• Pastes and viscous liquids flow slower than their card says along the ground: the bottom one or two particle layers move at about half speed.
• The flank of a collapsed heap reads lower than the 26-27 degree repose angle, and a squat collapse's runout still grows with resolution.
• Above 64 GPU worlds, sand and paste with a body in them are not validated (grains can sink into bodies).
• Sand and paste on a GPU use about 10x the memory with CUDA graphs; a scene too big for graphs runs without them, and one too big for both is refused.
• Sand, dough and paste at many worlds are limited by the material solve: 256 dough worlds take about 7.6 s per 10 ms period on an L4.
• On the dense grid, a world that outgrows the grid's box couples the worlds by one rounding step (1.5e-8 to 3.1e-8 m).
• Shear-thinning pastes thin a little too little: fast flow runs up to about 15 % slow.
• A material cannot both melt and shear-thin: the melt law is lost without a warning (no library material has both).
• A tool thinner than a paste's grid cell (whisk wires, skewers) drags it in pulses.

Health, determinism and coupling

• On a GPU, rigid bodies and robots in contact do not repeat bit for bit once a body takes more than one contact pair.
• Health's overlap flags (PEN_WARN, PEN_FAIL) for rigid bodies and robots do not yet catch every deep contact.
• Worlds of a coupled scene share one exchange clock, so one world's events change the steps of every world.
• A frozen INVALID world still sets a coupled scene's clock, so a start overlap can make each control step take minutes.
• With the opt-in contact gate, cloth resting on fixed geometry while the robot is away settles at a coarser step.
• Contact between a soft or cloth particle and a MuJoCo body takes its gains from the exchange clock.

Cameras

• Env camera images are bright and grainy: mean brightness 214 of 255 and 3.9 grey levels of noise at 4 samples per pixel.
• Glass objects render opaque unless a look file makes them glass.
• A robot file with one visual-only geom draws only its visual geoms, so the rest of the robot is invisible in images.