The rl_stl package lives in src/rl_stl/. It keeps simulation, synthesis dynamics, STL, synthesis, policies and the supervisor in separate modules. The JAX-native parts (MuJoCo Playground, MJX, Brax, the proposer and the differentiable check) never call the host-side Gurobi solver from inside a batched, JIT-compiled loop.
Module Overview
rl_stl.synthesis: the synthesis program and the check
| Module | Description |
istl.py | Receding fixed-horizon interval-STL synthesis (NAHS (16)) with Gurobi: RecedingProblem, History, persistent models per first window (RecedingISTL, ISTLCache) |
check.py | Independent check of a candidate plan against the receding I-STL constraints: check_plan, interval_robustness, precompiled windows for a fast runtime check |
terminal.py | Verified control-invariant terminal sets: polygon algebra, pre_set, max_control_invariant, verify_invariant, patrol_terminal_set |
tube.py | Error tubes: disturbance_tube for \(e^+ = (A + BK)e + Gw\), constant_tube for the NAHS-style constant box |
common.py | Shared Gurobi environment and the solver status enum |
| Module | Description |
receding.py | RecedingSupervisor: proposal, check, cached-plan execution and delay-compensated fallback. Plan holds the nominal plan with its root, error age and terminal face, extended past \(N\) by the terminal controller |
| Module | Description |
segment.py | Learned segment proposer: observation features, squash into the tightened input box, patrol_certificate (the differentiable margin \(c\), \(c \ge 0\) iff the plan passes the check), SegmentPolicy, LearnedProposer |
nominal.py | Nominal mission controller of the patrol: PD toward the current waypoint |
ppo_checkpoint.py | Config schema and policy loader for the X2 hover PPO runs |
rl_stl.stl: specifications and robustness
| Module | Description |
nahs.py | Patrol task: the square S, formulas (24) and (25), receding window specification |
robustness.py | Independent STL robustness evaluation and online (prefix) interval semantics |
jax_robustness.py | Compiles stlpy formulas into jittable JAX robustness functions, plus the capped hinge penalty |
specs.py | STL specifications built with stlpy |
rl_stl.dynamics: planning models \(f_{\mathrm{syn}}\)
| Module | Description |
linear.py | Discrete-time linear synthesis models (zero-order-hold double integrators) and helpers |
quadrotor_fsyn.py | Reduced-order synthesis model for the Skydio X2: 3-D double integrator |
rl_stl.envs: simulation models \(f_{\mathrm{sim}}\)
| Module | Description |
skydio_x2/model.py | Skydio X2 simulation model from MuJoCo Menagerie |
skydio_x2/ctbr.py | Normalized collective-thrust-and-body-rate action interface: rate loop with gyroscopic feedforward, X-configuration allocation, prioritized desaturation |
skydio_x2/tracker.py | Outer-loop tracker from a planning-model reference (position, velocity, acceleration) to a CTBR action |
skydio_x2/host_sim.py | Host-side MuJoCo X2 simulation driven through the CTBR interface |
skydio_x2/hover.py | MuJoCo Playground (MJX/JAX) hover task for the X2 |
skydio_x2/frames.py | Rotation and frame conventions |
blimp/model.py | MuJoCo blimp matched to the NAHS BlimpSim model, and its planning model |
patrol_plan.py | Planning-level patrol environment in pure JAX for PPO fine-tuning of the proposer |
| Module | Description |
systems.py | The three patrol benchmarks (double integrator, X2, blimp) behind one interface, and run_episode: the supervisor at the planning rate, the simulator below it |
patrol.py | Patrol experiment builders and the double-integrator closed loop |
energy.py | CPU energy measurement with Intel RAPL: EnergyMeter, benchmark |
| Module | Description |
config.py | YAML configuration files loaded into frozen dataclasses |
provenance.py | Run directories with resolved config, seeds, git revision, versions and hardware |
device.py | CPU/GPU switch for JAX, MJX and Brax |
plotting.py | Matplotlib defaults for publication figures (TrueType fonts, never Type 3) |
render.py | MuJoCo video helpers |
Generated Documentation
Doxygen also extracts the docstrings of every module, class and function. Use the search box, or start from a package: synthesis, supervisor, policies, stl, dynamics, envs, evaluation.