Native simulation frontier

Four source-backed native simulation workstreams. Newton and Genesis World are physics engines; gym-pybullet-drones provides aerial control environments; Stonefish provides marine simulation. These classes do not establish comparable benchmark scores.

Reviewed 14 September 2026. The initial feasibility review added zero worlds, recordings or reproduced evaluations. The subsequent standalone Newton FR3 qualification adds one native recording with independent control replay. The existing Isaac drawer recording uses Newton 1.5.0; it does not verify the standalone Newton 1.6.0 examples below.

Selected source inventory · Primary sources · Implementation sequence

Physics engines

Newton’s Panda hydro example includes a final lift check above 0.15 m and additional placement checks when a cup is selected. The qualified recording fixes the pen scene, one world, 720 frames and both deterministic options. Its asset downloader pins newton-assets to f8fb7abc and Menagerie to da76818e. The qualified FR3 recording separately inventories and audits its selected original assets; the other examples remain unqualified. The cartpole’s body-state assertions do not establish grasping or a learned-policy result.

Genesis’s grasp tutorial loads xml/franka_emika_panda/panda.xml. That model’s README traces it to franka_ros and states Apache-2.0. Licences in the separate panda_bullet and franka_sim directories are not evidence for this selected model. The tutorial supplies a sequence of commands without a scored task registration or explicit success predicate; lift and hold checks must be declared before recording.

Dependency declarations are not verified installations. A separate environment preserves the working Isaac backend. See the pinned Newton and Genesis dependency files. Genesis World is the open-source engine in the software catalogue, distinct from Genesis AI’s commercial products.

Aerial control environments

The PID example is a controller demonstration. CtrlAviary returns reward −1 and false termination/truncation flags. Reaching the loop horizon does not establish success. A single-drone or obstacle-free trial must be labeled as an authored variant.

The repository’s code licence is MIT. The selected COLLADA model names Lakshman Kumar in its author metadata; that attribution and the separate PyBullet asset notices must be traced in the native input inventory. No model geometry is hosted by this review. The installation notes and dependency constraints describe source requirements, not a runtime tested here.

Marine simulation

The example manager contains both PARSED_SCENARIO and programmatic scene branches; the compiled choice must be recorded. The parsed scenario includes an ocean current, three Girona vehicles and spline animation. Its existence does not establish vehicle control. The example starts at 200 Hz and its step-completed callback supplies no mission success predicate. Native NED coordinates, sensor rates and actual thruster setpoints must remain explicit.

The installation guide allows a local build prefix; CMake defines the build dependencies. The native trial must bind the scenario’s included files and original model/media references before capture.

Implementation sequence

This is the proposed order within this frontier workstream. Newton’s first recording is qualified; the remaining targets require their own native qualification. The linked tracker is private and requires repository access; the primary sources and this review are public.

1. Newton follow-up #245: completed for one native FR3 pen/cup recording, including independent replay; see the qualified evidence.
2. Genesis follow-up #246: bind the exact Panda model and dependencies, declare lift/hold checks, then qualify the grasp tutorial.
3. Aerial follow-up #247: resolve the current Python/PyBullet build, declare one bounded PID trajectory and measure its tracking and collision outcomes.
4. Marine follow-up #248: verify the selected Stonefish build/scenario, add a bounded controller protocol, then measure the underwater trial.

For each: bind sources, original assets and separate terms; declare initial conditions and outcomes; record native camera/state/control channels; replay controls independently; check media and desktop playback; review, merge and verify production. Original geometry and physics stay native. A recorded example can establish a bounded simulation result; it cannot establish a benchmark suite score.

Review coverage

The inventory records 39 selected files, their pinned URLs, Git blob identities, sizes and SHA-256 hashes. Downloaded bytes matched the corresponding regular-file entries in the four non-truncated Git trees. This is a selected source review, not a complete dependency or asset lock. No upstream code was executed for that initial feasibility inventory. The later Newton qualification has its own complete input and runtime evidence linked above.