---

name: rl-mpc-locomotion description: "Guide quadruped locomotion workflows built from Python convex MPC, FSM and gait control, NVIDIA Isaac Gym simulation, Hydra/RSL-RL training, learned MPC-weight policies, robot assets, and CUDA or solver troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: MIT

RL-MPC Locomotion

Use this repo skill when a task involves the rl_mpc_locomotion package, its MPC_Controller and RL_Environment modules, quadruped torque control, Isaac Gym locomotion simulation, or learned MPC-weight policies. This is a router, not a replacement for the focused workflow references.

Gate before acting

1. Read repository provenance when checking whether the guidance matches the source revision.
2. Use setup-and-diagnostics first for an isolated Python 3.8-era environment, pinned native inputs, mpc_osqp, CUDA, Isaac Gym, asset, and checkpoint probes.
3. Treat Isaac Gym Preview 4 as a required external backend for simulation, RL training/evaluation, and the interactive launcher. The construction evidence verified PyTorch 1.10/CUDA 11.3, CUDA allocation, repository imports, rsl_rl, and mpc_osqp, but did not verify Isaac Gym. Do not substitute a CUDA tensor or CPU import for that gate.

Route by task

• MPC, gait, FSM, state estimation, robot model, torque arrays, or solver binding: read mpc-control.
• Hydra training, evaluation, task YAML, observations/actions/rewards, checkpoints, policy bridge, or TensorBoard: read rl-training, then route low-level torque questions to mpc-control.
• Isaac Gym setup, PhysX, URDF/assets, terrains, state tensors, effort control, viewer lifecycle, or gamepad launcher: read isaac-gym-simulation.
• Installation order, submodules, PyTorch/CUDA conflicts, compiled extension, optional solver, missing config/asset, or backend diagnosis: read setup-and-diagnostics.

For shared concepts and cross-route boundaries, read the architecture overview. For a cross-cutting failure, read troubleshooting.

Minimal public package check

In a current project copy, install the package after its documented native inputs are present:

python -m pip install -e .
python -c "import MPC_Controller, RL_Environment, mpc_osqp; print('MPC package imports passed')"

Run the bundled checks from the generated skill root (or enter the setup sub-skill directory). Set PROJECT_COPY to the current user-supplied project copy when validating its native layout/config/assets; it is not a path inside this skill bundle. The checks are read-only unless --build is passed explicitly to the extension checker:

PROJECT_COPY=/path/to/current/project-copy
python sub-skills/setup-and-diagnostics/scripts/check_environment.py --pip-check
python sub-skills/setup-and-diagnostics/scripts/check_mpc_extension.py --repo-root "$PROJECT_COPY" --check-submodules --strict
python sub-skills/setup-and-diagnostics/scripts/validate_config.py --repo-root "$PROJECT_COPY" --all-tasks --strict

The setup sub-skill documents the explicit project-root argument needed when checking a particular checkout. These commands do not prove Isaac Gym or a stable gait.

Safety and handoff

Never start a viewer, gamepad reader, training run, or policy deployment as an installation smoke test. Confirm robot/task, device, checkpoint, run directory, backend availability, and cleanup behavior first. The project supports Aliengo, A1, and Go1 in the public runtime selection; the bundled Mini Cheetah assets are not a launcher robot choice. Sim-to-real operation is outside this repository's verified contract.