Modelling Envelope

Simulation results do not establish physical accuracy or hardware readiness. See the changelog for corrections and the capability ledger for current evidence status.

Dynamics

Orbital tasks are microgravity, not orbital mechanics

ORBIT uses zero gravity; SRB does not model orbital rates or Clohessy-Wiltshire / Hill relative motion. Tasks assess contact and manipulation in microgravity, not rendezvous geometry, formation flight, or station-keeping fuel use.

Manipulator arms do not feel domain gravity

Manipulators, grippers, and tools have gravity disabled, regardless of env.domain; results do not compare performance across planetary gravities. The Raph shovel lifts but does not curl; see the hardware shovel decision.

Aerial tasks are kinematic, not aerodynamic

Multicopter actions set root velocity and attitude; animated rotors produce no thrust. Tasks assess navigation and guidance, not aerodynamics, disturbance rejection, power, or endurance.

Large spacecraft inertia comes from a uniform bulk density

Spacecraft without explicit mass use uniform bulk density, not hollow-structure inertia; they do not support attitude-control conclusions. Landers use explicit mass and per-thruster forces. mobile/landing is powered descent, not a hardware-fidelity claim.

Terrain and contact

There is no terramechanics, and friction is profile-specific

Ground contact uses Coulomb friction on rigid surfaces, without sinkage, slip-sinkage, bulldozing resistance, or drawbar pull. Profile friction values are engineering parameters, not physical calibration; results do not establish regolith traction or trafficability.

Self-collision is disabled on most articulations

Many articulations do not test self-collision. Configurations may be physically unreachable; check collisions before transferring a policy.

Actuators are implicit PD

Most actuators use implicit PD without torque-speed curves, joint friction, backlash, compliance, current limits, or thermal limits. Results do not establish actuator or torque-tracking fidelity.

Granular material

Regolith has no cohesion, and cannot be given any

The shipped non-fluid PBD model ignores cohesion and adhesion settings. It cannot represent cohesive regolith or predict excavation force, scoop fill, trench walls, bootprints, or tracks.

Particles are two to three orders of magnitude too large

Excavation defaults to 7 mm particles and landscaping uses 5 cm particles, versus lunar grains around 70 µm. These scales omit finer interactions and features.

Real terrain

Measured terrain resolves at source-specific native GSD; finer detail is synthetic

Measured terrain resolves to each source’s native ground sample distance; resampling adds no detail, and smaller-scale roughness is procedural. See the real terrain guide for source and cache details. Terrain shape at source resolution is not evidence of wheel-scale traversability.

Seven of eleven named sites still carry placeholder sources

Some named sites have placeholder source digests and fail closed before download. Check the terrain catalog; a declared digest alone is not execution or fidelity evidence.

Named-site fusion covers one declared Jezero stack

Named-site multi-source fusion covers only the declared Jezero stack. See the terrain stack guide for current coverage.

Sensing

Noise and delay lack target-hardware calibration

Noise and delay are not calibrated against target hardware. The optional waypoint_navigation_synthetic_v1 profile is not sim-to-real validation. See randomization profiles.

Depth noise is modelled on consumer hardware

Depth-noise parameters are based on consumer Intel RealSense devices; calibrate against the target device before transfer.

Illumination and appearance

Light intensity is documented in W/m² but consumed as a renderer scale

Light intensity is a renderer scale, not calibrated irradiance. Images do not support absolute photometric or radiometric claims. With env.ephemeris, SRB can provide relative sun direction.

Ambient sky fill is one undifferentiated constant

Ambient fill is constant across domains, not derived from a measured sky.

Mars sunlight is not reddened

Mars dust reddening is not modelled; rendered color is not calibrated.

The default lunar surface material is the wrong colour

Under the solar illuminant, the procedural lunar material is magenta, not lunar grey (R 122.4 / G 65.9 / B 138.7; G/R = 0.54). Vision results are not color-faithful; geometry, light direction, and shadows are unaffected.

The Earth domain renders a starfield sky

The Earth domain uses a starfield, not a daylight laboratory background.

Time discretisation

Aerial tasks run at most two physics steps per control step

Aerial tasks use one or two physics steps per control step; conclusions about dynamics near or above the control rate are unsupported.

GPU contact-buffer sizing is on Isaac Lab’s defaults

SRB uses Isaac Lab’s default GPU contact-buffer sizing. Check large CUDA runs for PhysX overflow warnings.

Domain constants

Asteroid gravity randomisation reaches zero

Asteroid gravity randomization can reach zero.

Two of the five domains are the default of no task

ASTEROID and EARTH are selectable, but no shipped task defaults to them and no task acceptance path has exercised them. Selecting a domain is not validation.

Training configuration

Two of the 42 tasks have tuned hyperparameters

Most task and algorithm combinations use defaults; scores describe those configurations, not best performance or controlled algorithm comparisons.

Verification

Physical assertions are the exception, not the rule

Tests primarily check software contracts; passing them does not establish physical accuracy.

Release validation fails on three landscaping tasks and three event-camera variants

Terrain-landscaping and some event-camera paths have known runtime limits. See the terrain landscaping workflow and event-camera status. Synthetic smoke tests do not establish real-data fidelity.

Known software defects

The defects below are declared so this release does not imply support it lacks. Corrected defects are removed from this list and recorded in the changelog.

Terrain conversion and resampling

• PDS3 conversion drops calibration. srb.terrain.sources.pds_img.img_to_geotiff writes raw stored samples without the label’s SCALING_FACTOR and OFFSET. No catalog bake in this release calls it; do not treat its output as calibrated.
• The optional pdr PDS3 fallback drops georeferencing and nodata. It writes no CRS, an identity transform, and no nodata. Supply the correct georeferencing and nodata yourself before using its output.
• An explicitly supplied orthophoto is not spatially aligned to the DEM. ortho_path images are resized to the DEM grid by array shape alone; extent and transform are ignored. Images not pixel-aligned with the DEM stretch onto the wrong ground. The fused v2 colour path is aligned.
• Decimated reads do not guarantee averaging. Reprojected sources are warped onto the planned output grid with nearest-neighbour sampling, and a direct read applies no explicit downsampling filter. Averaging therefore depends on source overviews, and fine elevation detail can alias.
• Material baking has no aggregate memory budget. Per-face and per-block limits apply, but total intermediate allocation across all faces is unbounded.

Restrictions: Use only an orthophoto pixel-aligned with its DEM, and do not rely on decimated reads to preserve band-limited elevation detail.

Learning configuration

• Shipped Robomimic bcq and iris profiles fail training preflight because they disable components those algorithms require.
• For agents with dictionary configurations, an agent mapping in a --cfg file replaces the registered mapping instead of merging. Include every required agent setting in the file.
• The excavation and spacewalk blocks in hyperparams/dreamerv3.yaml and the excavation block in hyperparams/tdmpc2.yaml are not read. Dreamer task settings that apply live under hyperparams/task/<task>/.
• Dreamer real transfer saves checkpoints only at its periodic interval; a normal stop loses progress since the last save.

Event-camera provenance and capture

• Two event sensors in one environment with the same base seed receive the same derived stochastic seeds and hot-pixel mask seed; sensor identity is not included. Whether emitted noise is correlated has not been measured.
• The EVK4 capture adapter refuses a second SDK callback before the pending packet is polled. Capture therefore fails closed if callbacks arrive in bursts. This has not been checked against the supported OpenEB SDK and device.

Model artifacts and shared caches

Model and cache operations validate their inputs and can refuse incomplete or inconsistent artifacts. See the model reference and cache commands.

Real-robot interface

• SpaceMouse motion is discarded after no device read succeeds for 0.5 s. Reads can still succeed while reports are lost, leaving the last motion active. Do not drive a robot without an independent deadman.
• Dreamer real-robot startup resets the environment in both the CLI and Dreamer’s warm-up; each reset may resume eligible command sinks. Whether both resets are intended is unresolved.
• The draft hyperparams/validation/spacewalk.yaml treats distance and wall-time metrics as higher-is-better. It is not runnable in this release; set metric directions before enabling it.

Restriction: This interface does not support hardware operation; see Sim-to-real. Generated direct-input deployment profiles also remain unavailable.

Installers

• If scripts/install_isaaclab.bash, or the Dreamer, RSL-RL, or Robomimic installer fails after installing into the Isaac Sim interpreter, it keeps the checkout because editable installs may point into it, and prints retry instructions. It does not roll back packages pip already installed or upgraded. A rerun reinstalls editable packages (the Isaac Lab installer needs a new destination), but packages from the failed run remain; recreate the Isaac Sim Python environment to remove them.

Reporting a gap in this page

Report missing or incorrect modelling limitations through GitHub Issues.