feat: expand the multi-robot simulation stack

[Functionhx]

What changed

• migrate and validate the Isaac Lab backend on Isaac Sim 6.0.1 GA, including robust/stair training, transfer tooling, and audited capsule-compatible assets
• add shared 1/2/3-link inverted-pendulum tasks across MuJoCo and Isaac Lab, with PPO, SAC, TD3, classical/robust/observer/trajectory controllers, and same-contract evaluation
• add the serial wheel-legged URDF/MJCF pair, Isaac Lab training/evaluation/export, and MuJoCo sim2sim tooling
• add the 11-DOF ActuateX Sentinel as a paired MuJoCo/Isaac Lab system: shared 28-D observation and 6-D action contract, 500 Hz explicit DC-motor physics, power-buffer/thermal state, launcher safety and ballistics, PPO/SAC/TD3 training, and bidirectional evaluation
• add native local system identification in MuJoCo and Isaac Lab, command-space gain scheduling, LQR synthesis, and a genuine discrete-time H-infinity zero-sum-game Riccati solver with saddle-Hessian and bounded-real frequency certificates
• add a shared six-segment remote-command protocol, continuous CSV/SVG traces, fall/power/thermal/buffer telemetry, native evaluation, and bidirectional sim2sim classification behind one policy-loader contract
• add a ROS 2/Nav2 integration package with camera profiles, MID360 non-repetitive point timing, point-cloud conversion, maps, and launch/config files
• add the first industrial humanoid layer: a Unitree G1 29-DOF SDK-order contract with 50 Hz timing, 480-D term-major observation history, motor torque-speed/friction envelopes, action latency, and fail-closed runtime guards
• refresh bilingual READMEs, architecture/migration/benchmark reports, the code-change PDF, third-party provenance, and repository verification
• document a primary-source RoboMaster community mechanical survey, the current Sentinel model, Isaac Sim fidelity limits, licensing boundaries, and the next identification/training milestones

Why

ActuateX is growing from a single quadruped coursework baseline into a reviewable multi-robot, multi-simulator control stack. The contract layer makes simulator-to-simulator and future simulator-to-real claims testable by freezing joint order, frames, observation layout, actuator semantics, timing, and acceptance metrics.

The Sentinel phase now closes both the dynamics loop and the classical-control loop across two independent physics engines. Local models are rejected unless their validation, equilibrium, rank, controllability, and fall gates pass. Scheduled LQR and H-infinity artifacts carry source hashes and backend metadata, so native and cross-engine results cannot be silently mixed.

The H-infinity implementation is not a relabeled LQR. It solves an indefinite discrete zero-sum game, verifies the control/disturbance saddle conditions, checks Schur stability, and certifies the disturbance-to-performance gain on a dense unit-circle frequency grid. Runtime tests also preserve negative evidence: local linear certificates do not guarantee robustness to contact changes, actuator saturation, large command reversals, or broad domain randomization.

The Sentinel asset remains an original serial-leg design. Community CAD is not redistributed, and restricted mechanical releases are referenced only as design and failure-mode research. The model is control-grade nominal physics, not yet a team-specific digital twin.

The G1 and Sentinel smoke runs prove that the local Isaac/PhysX and MuJoCo pipelines launch, sample, update, and exchange policies. They are intentionally not presented as completed humanoid or competition-grade wheel-legged training. Long-running multi-seed RL training, terrain/impact/self-righting acceptance, and parameter identification remain follow-up work.

Validation

• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 pytest -q backends/mujoco/tests/test_sentinel_linear_control.py backends/mujoco/tests/test_sentinel_evaluation_protocol.py backends/mujoco/tests/test_sentinel_evaluation_trace.py backends/mujoco/tests/test_robomaster_dynamics.py — 32 passed
• H-infinity unit gate — gain differs from LQR; saddle Hessian signs, Schur stability, and frequency norm below gamma all passed
• Ruff on every changed Python file — passed
• python -m compileall -q ... — passed
• git diff --check — passed
• clean full-sequence scheduled LQR, 64 environments and zero falls in all four directions: Isaac native 0.0587 RMSE; Isaac to MuJoCo 0.0616; MuJoCo native 0.0417; MuJoCo to Isaac 0.0544
• clean full-sequence scheduled H-infinity, 64 environments and zero falls in all four directions: Isaac native 0.0619 RMSE; Isaac to MuJoCo 0.0718; tuned MuJoCo native 0.0409; MuJoCo to Isaac 0.0536
• Isaac training-randomization and held-out stress modes exercised mass/COM, friction, actuator, delay, reset, push, contact, power, buffer, and thermal limits; results retain the observed PPO advantage over local linear control under broad nonlinear perturbations
• Sentinel MuJoCo asset compile — nq=18, nv=17, nu=11, 44 sensors
• Isaac Sim 6 Sentinel rigid-body smoke — 4 environments x 250 physics steps
• Sentinel PPO smoke — MuJoCo 8 environments x 1 iteration; Isaac Lab 4 environments x 1 iteration
• Sentinel SAC and TD3 smoke — MuJoCo 8 environments / Isaac Lab 16 environments, 1024 transitions each, with GPU updates
• bidirectional inference smoke — MuJoCo SAC to Isaac Sim and Isaac SAC to MuJoCo both completed with shared observations, actions, motor temperature, chassis power, and referee-buffer telemetry
• MID360 .xz integrity and 800,000-row count — passed
• code-change PDF rebuilt successfully (12 A4 pages)

Generated logs, checkpoints, identification models, evaluation JSON/CSV/SVG, videos, simulator installations, and external Unitree/Isaac repositories remain excluded from Git.