Modelling Envelope

1. Dynamics

Orbital tasks are microgravity, not orbital mechanics

ORBIT sets gravity magnitude and variation to 0.0. It supports zero-gravity contact, manipulation, and free-flyer tasks.

It omits Clohessy-Wiltshire/Hill relative motion and orbital rate. It cannot support formation flight, rendezvous geometry, or station-keeping fuel budgets.

Manipulator arms do not feel domain gravity

Twenty robot/object files contain 24 disable_gravity=True declarations across manipulators, grippers, tools, floating, and aerial robots.

PhysX gravity is correct to five decimals in all five domains, but affects only free bodies and mobile bases. MOON, MARS, and EARTH change lighting and unattached objects, not arm loads. The excavation shovel is weightless.

Raph’s hardware shovel has one lift joint and no curl joint; it only raises or lowers the blade. These assumptions suit orbital arms, not surface arms. Do not claim cross-domain gravity transfer.

Aerial tasks are kinematic, not aerodynamic

Aerial tasks support navigation and guidance above flight control.

The multicopter term writes root velocity and matching attitude. Nominal-RPM rotors are visual and produce no thrust. There is no aerodynamics, rotor inflow, blade-element model, or airframe gravity. It follows body-frame velocity commands with bounded tilt.

Do not claim flight dynamics, control authority, disturbance or wind response, power, or endurance.

Large spacecraft inertia comes from a uniform bulk density

Large spacecraft without authored mass use uniform 1500 kg/m³ density, several times higher than mostly hollow spacecraft. Scaling corrects mass, not inertia: uniform solids differ from wall-loaded structures, and inertia ratios govern attitude dynamics.

Landers have explicit mass, enabled gravity, and per-thruster body forces via ThrustAction. Do not use large uniform-density spacecraft for attitude-control results.

mobile/landing supports powered descent.

2. Terrain and contact

There is no terramechanics, and friction is profile-specific

Use wheel-ground contact for navigation, path following, and obstacle negotiation. Do not use it for slip-ratio/sinkage predictions, traction budgets, or regolith trafficability.

BaseEnvCfg defaults to 0.50/0.50 Coulomb friction (static/dynamic), 0.0 restitution, and average combine mode. Wheels use actuator-driven velocity/position targets on rigid surfaces. There are no sinkage, pressure-sinkage, slip-sinkage coupling, bulldozing resistance, drawbar pull, or Bekker/Wong models.

The baked terrain path can apply these unvalidated engineering estimates. The registry has no asteroid entry.

Without baked material, the global default is a parameter, not calibration. Pre-materials-v3 bakes lack bindings.

To restore bindings:

• Re-bake with srb asset download terrain <site>.
• For a region, add --body, --lat, --lon, and --size-m to that command.
• Alternatively, enable auto-fetch.

With SRB_TERRAIN_AUTO_FETCH=0, direct terrain construction raises TerrainNotBakedError for unbaked terrain; there is no old-bake fallback. Runtime surface selection instead warns and falls back to procedural terrain. Isaac Sim contact effects are unvalidated; asset materials contribute through average combine.

Self-collision is disabled on most articulations

Self-collision is off on roughly 70% of articulations, so robots may pass through themselves in hardware-unreachable configurations. Require an independent collision check before claiming whole-body transfer.

Actuators are implicit PD

Many actuators use implicit PD; some configurations include joint friction. Most omit backlash, gear compliance, current, and thermal limits. Do not claim torque-tracking fidelity or results dependent on actuator dynamics rather than kinematics and contact.

3. Granular material

Cohesion and adhesion settings are ineffective for the current non-fluid particle setup

Excavation particles set cohesion and adhesion to 0.0. Landscaping sets cohesion to 0.0 and inherits adhesion 0.05; other scenes inherit base values 0.01/0.05.

PhysX applies cohesion and adhesion only to fluids, so they do not affect non-fluid PBD particles; friction does. The angle of repose is uncalibrated against lunar regolith and simulants. Without cohesion, particles cannot form bootprints or trench walls or hold scoop loads; another representation is required.

Do not claim excavation forces, scoop fill, trench-wall stability, bootprints/tracks, or simulant comparisons.

Particles are two to three orders of magnitude too large

Excavation particles are about 0.7 cm; landscaping particles are 5 cm. Lunar regolith’s median grain size is about 70 µm (Carrier et al., Lunar Sourcebook, 1991), requiring about 10¹² particles per cubic metre.

Coarsening is unavoidable. The model lacks grain interlocking, fines-scale cohesion, and sub-particle features. Do not claim results below particle diameter, including tool-edge detail or fine texture.

4. Real terrain

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

Use measured terrain for site-specific slope, illumination, and route-scale shape at source GSD. Do not use it for wheel-scale traversability. The real-data path is not fully validated.

Its pipeline records curated, SHA-256-pinned LROC NAC DTMs, IAU_Moon_2015 CRS, per-site GSD, licence, attribution, and SPICE sun-position epochs. Native GSD varies by source:

Resampling and fusion add no detail below source GSD. Procedural rocks, craters, and noise supply synthetic rover-scale roughness, even at real sites. See Terrain Stacks for site and multi-source stacks.

5. Sensing

Noise and delay lack target-hardware calibration

Default task configs leave Isaac Lab’s observation_noise_model and action_noise_model unset. Waypoint navigation retains relative observation perturbations. Opt-in waypoint_navigation_synthetic_v1 adds Gaussian action noise and (1, 2)-step action/observation delays.

See randomization profiles for parameters and reset behavior. Noise lacks hardware calibration and realistic IMU/image models. Do not claim target-device robustness without measurements.

Depth noise is modelled on consumer hardware

Heightmap and depth models include quadratic stereo and time-of-flight noise, quantisation, incidence-angle and random dropout. Parameters come from consumer Intel RealSense devices; space-qualified sensors may differ.

Calibrate and re-parameterise for the target device before claiming sensor-specific transfer.

6. Illumination and appearance

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

Use configured sun elevation, azimuth, and shadow direction; SPICE supplies these when env.ephemeris is enabled. Uncalibrated rendered radiance gives no absolute photometry or radiometry.

Domain.light_intensity is documented in W/m², but USD DistantLightCfg consumes renderer scale, not irradiance. Nominal intensities are 1361 W/m² for Moon/orbit and 775 W/m² for Earth; Earth’s range reaches 1000 W/m² (ASTM G173 AM1.5 clear-sky surface value). Mars’s 590 W/m² is top-of-atmosphere, despite the docstring; SRB omits typical surface irradiance (~60–70%, dust-dependent).

Ambient sky fill is one undifferentiated constant

All domains, including airless ones, apply DomeLight at 25% of direct intensity, without a derivation for 0.25. It barely affects Moon shadows because stars.exr is near-black. Changing the constant or texture could sharply alter lighting.

Mars sunlight is not reddened

Iron-oxide dust reddens Martian sunlight, producing a butterscotch appearance. SRB uses the Sun’s 5778 K colour temperature and has no calibrated reddening value. Do not claim colour-accurate Mars imagery.

The default lunar surface material is the wrong colour

Upstream simforge_foundry owns the installed MoonSurface material. It renders red-brown to magenta under solar illumination, unlike lunar regolith’s grey with a slight brown cast. Geometry, illumination direction, and shadows are unaffected. Do not use default-path datasets for colour-dependent perception on procedural lunar terrain.

The Earth domain renders a starfield sky

Domain.EARTH uses stars.exr as its skydome, giving daytime scenes a night sky; daylight textures exist in the asset pack. Do not transfer Earth-domain background vision models to real laboratories.

7. Time discretisation

Aerial tasks run at most two physics steps per control step

Do not draw aerial-dynamics conclusions near control-rate frequencies.

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

SRB applies no per-num_envs GPU buffer sizing; PhysX buffers keep Isaac Lab’s defaults. CPU defaults are unaffected; check PhysX overflow warnings on CUDA runs with large num_envs.

8. Domain constants

Asteroid gravity randomisation reaches zero

ASTEROID gravity variation is twice its magnitude, giving a sampled lower bound of 0.0 m/s². Some episodes run in free fall.

No task defaults to the Asteroid domain

srb/core/domain.py declares five domains. Task configs explicitly default to MOON (6), MARS (4), and ORBIT (2), with none for ASTEROID; others inherit their family default. modular_assembly keeps the default domain but sets gravity to EARTH (srb/tasks/manipulation/modular_assembly/task.py:170). Do not treat an ASTEROID switch as validated.

9. Training configuration

Two of the 42 tasks have tuned hyperparameters

hyperparams/ has 28 reinforcement-learning and 11 imitation-learning configs; hyperparams/task/ overrides two of 42 tasks. The other 40 use defaults, not best-achievable settings. Untuned scores are unfair across algorithms; published scores are not tuned or configuration-matched.

Known limitations