Call GRiD from Python (numpy / JAX / torch) |
grid_rbd.load_robot("robot.urdf", backend=...): one call, no name
needed. Guided tour: Python Wrappers (grid-rbd);
API: grid_rbd (Python bindings); GPU-resident lifecycle notes
(stay on device, CUDA graphs, zero-copy interop):
bindings/examples/AGENT_INTEGRATION_GUIDE.md.
|
Generate CUDA for a new robot |
grid-generate path/to/robot.urdf [-f] (ten sample URDFs in
grid_assets/robots/); walkthrough:
Generate CUDA for a robot.
|
Make a big (humanoid) robot build fit in RAM or finish faster |
Fast Robot Setup: subset the build
with algorithm_list= and/or skip the mjx twins with
enable_mujoco_kernels=False (also grid-generate
--algorithm-list ... --no-mujoco-kernels). |
Use my own top-level GLASS instead of the copy vendored in grid.cuh |
gen_all_code(..., vendor_glass=False): the header #includes
glass.cuh from your include path (-I<GLASS root>) and aliases
grid::glass to ::glass, one GLASS per translation unit. The
default (vendored, self-contained) header is byte-identical. All
*_DYNAMIC_SHARED_MEM_BYTES<T[, TIER]>() sizers are constexpr.
|
Label the GLASS revision when generating from a source archive (no .git) |
gen_all_code(..., glass_revision="<sha>") or
GRID_GLASS_REVISION=<sha>: the // Pinned commit: line carries
the bare revision, so the header is byte-identical to a git checkout’s;
a live checkout that disagrees is an error, and
gen.glass_revision_source reports git / git-verified /
supplied-unverified / unknown. Distribution artifacts record
their packaged pin automatically and report bundled.
|
Embed the generated header in a library or interpreter (no exit() on CUDA errors) |
init_robotModel_checked / init_joint_limits_checked /
free_robotModel_checked (or robotModel_owner<T>):
Library-safe initialization and cleanup.
|
Add a new algorithm to GRiD |
Adding a New Algorithm (numpy oracle in
RBDReference first, then the codegen emitter, then equivalence
tests). |
Run / verify the test suites, or fix a red receipt-verify CI job |
CUDA Validation And Performance Reporting: the marker map, the
split-suite driver, and the two-tier gpu-proof.json receipt policy
(a red verify job after touching fingerprinted tests is expected; run
the everyday refresh and commit the receipt). |
Benchmark GRiD (or compare against Pinocchio / MJX / Warp) |
Benchmarks. |
Debug a CUDA-vs-numpy mismatch or an unexplained kernel failure |
the debugging guide
(docs/agent_debugging_guide.md): the catalogue of known bug
classes (shared-memory init, output-convention errors, reduction
nondeterminism, launch-config pitfalls, …). |
Get MuJoCo/mjx-convention inputs and outputs |
handle.mujoco.<method>(...) (per-call, thread-safe) or
output_convention="mujoco": values and derivative/second-order
surfaces, floating base; see the conventions section of
Python Wrappers (grid-rbd).
|
Run several pipelines on one GPU without them sharing scratch |
handle.context() (numpy, jax and torch handles alike) opens an
isolated runtime context on the same artifact (own arena, tables,
streams, launch overrides), closed with that handle;
handle.device_profile says what it was fitted to. One pipeline per
context. Runtime contexts.
|
Swap inertias / attach a tool at run time and keep autograd consistent |
set_inertia_params / attach_tool / set_joint_dynamics
mutate the context under an exclusive admission lock and bump
handle.model_version; a torch/JAX backward whose forward ran under
an older version raises, so recompute the forward. Same page as above.
|
Know what this release supports, what changed, and what it does not do |
Compatibility and known limitations (platforms, contexts,
graphs, operands, native-interface stability, differentiation, build cost). |
Know what shapes / dtypes / devices a call accepts (and rejects) |
One rule set per operand class, enforced natively on every surface:
Operand validation. |
Understand why something recompiled (or refused to) |
The per-robot .so cache key: Fast Robot Setup;
the test suites’ compile caches and receipt fingerprints:
CUDA Validation And Performance Reporting. |
See what a registration would build, or warm the cache out-of-band |
grid_rbd.build_plan / grid_rbd.precompile:
Fast Robot Setup.
|
Tune kernel launch configs for my GPU |
grid_assets/autotune_robot.sh --help (writes
grid_assets/launch_configs/<robot>/<gpu>.json, baked at codegen time,
overlay-able at runtime via apply_profile_overlay).
|
