Tutorials

This section has two tracks: using GRiD and contributing to GRiD.

Using GRiD

• Python Wrappers (grid-rbd)
  • Install
  • Register-then-run UX
  • Method surface
  • Precision
  • Build cost on large floating-base robots
  • Cache layout
  • Dimensions, layouts and differentiability
  • End-effector target selection
  • Contact frames and welded tools
  • Host round trips (numpy): allocate once, reuse
  • JAX FFI (grid_rbd[jax])
  • PyTorch backend (backend="torch")
  • grid_plant cost / barrier / plant-step methods
  • External forces (f_ext)
  • See also
• Python backend interfaces
  • How to interpret availability
  • Joint and interface restrictions
• CPU-checked input examples
  • Fixed and floating input packing
  • Constructing a regressor parameter vector
  • An integrator diagnostic
• Generate CUDA for a robot
  • Generate a header
  • Profiles
  • Generate from Python
  • Use the generated code
  • Include it in your kernel
  • Compile time and footprint
• Fuse dynamics into your own kernel
  • The control step
  • Generate the header
  • Write the kernel
  • Initialise and launch
  • The layered path
  • Why fusion is faster
  • Build, check and time
  • Adapting the kernel
• Gradient descent through GRiD with plain JAX
  • What GRiD provides and what JAX provides
  • Load the robot
  • Targets
  • Loss and gradient
  • The descent as one GPU program
  • Run it
• Collision code generation
  • Generate collision routines
  • Device API surfaces
  • Geometry coverage and correctness
  • Sphere models and fidelity reports
• RBDReference
  • Run a reference calculation
  • Methods and validation
  • Dependencies
• Inspecting a robot model with URDFParser
  • Parse and look around
  • Checking a model before generating code
• CUDA Support Status
  • Joint and interface restrictions
  • Fixed-Base Robots
  • Floating-Base Robots
  • Shared-Memory Fallbacks
  • Known Caveats
  • Algorithm catalog

Contributing to GRiD

• Adding a New Algorithm
  • Generator functions
  • Step-by-step recipe (worked example: fdsva_so)
  • Common pitfalls
  • Code-generation helpers
  • Submitting the algorithm
• CUDA Validation And Performance Reporting
  • The launcher
  • The driver and the thin wrapper
  • GPU-Proof Signed Receipts
  • The two-tier receipt policy
  • CUDA Artifact Cache
  • Progress Output
  • Sample Selection
  • Shared-Memory And L2 Controls
  • Register-Pressure And Tier Analysis
  • Linear Algebra Backend
  • Performance Reporting
• Benchmarks
  • Quick start
  • Algorithms measured
  • Output schema
  • Stability and compile flags
  • Per-host autotuning (recommended for production)
  • Autotune launch config for your robot / GPU
  • Re-baking launch configs: refresh_launch_configs.sh
  • Pre-GLASS regression check
  • GPU-resident pipelines (no-transfer timing)
  • Second-order derivatives: transfer cost
  • Staying resident (jax)
  • See also