mjwarp_adtesting

Autodifferentiation testing suite for mujoco_warp. Validates AD correctness, performance, and visual fidelity across three dimensions:

• Mathematical correctness: AD vs finite-difference gradient comparison, Taylor convergence, chain rule verification
• Performance overhead: Wall-clock timing, GPU memory, scaling with DOFs and batch size
• Visual verification: Trajectory parity (AD on/off), gradient sensitivity heatmaps

Setup

# Install in editable mode with all extras
pip install -e ".[visual,analysis]"

Requires mujoco_warp installed or available at C:\Projects\mujoco_warp (configurable via --mujoco-warp-root or MUJOCO_WARP_ROOT).

Usage

# Run all non-slow tests
pytest mjwarp_adtest/ -v -m "not slow"

# Run by category
pytest mjwarp_adtest/math_tests/ -v -m math
pytest mjwarp_adtest/perf_tests/ -v -m perf --save-results
pytest mjwarp_adtest/visual_tests/ -v -m visual --save-results

# Skip tests requiring SM 70+ GPU
pytest mjwarp_adtest/ -v -m "not gpu_sm70"

# Run on CPU
pytest mjwarp_adtest/ -v --cpu

# Generate report from saved results
python scripts/generate_report.py --results-dir results

Test Categories

Marker	Description
math	Mathematical correctness (AD vs FD, Taylor, chain rule)
perf	Performance benchmarking (overhead, memory, scaling)
visual	Visual verification (trajectory parity, gradient viz)
slow	Long-running tests (humanoid, batch scaling)
gpu_sm70	Requires CUDA compute capability 7.0+

Configuration

Command-line options:

Option	Default	Description
--fd-eps	1e-3	Finite-difference perturbation
--fd-tol	1e-3	AD vs FD tolerance
--contact-fd-tol	1e-2	Tolerance for contact tests
--results-dir	results	Output directory
--save-results	false	Save results to files
--mujoco-warp-root	C:/Projects/mujoco_warp	mujoco_warp source root

Environment variable overrides: ADTEST_FD_EPS, ADTEST_FD_TOL, ADTEST_CONTACT_FD_TOL, MUJOCO_WARP_ROOT, ADTEST_RESULTS_DIR.