BNB-QP Reach-Avoid Solver

This repository contains a JAX-based branch-and-bound QP solver for STL-style reach-avoid trajectory optimization. The branch-and-bound layer fixes binary logic variables, and each continuous node relaxation is solved by the structured ADMM / TVLQR backend in primal_dual_ilqr/.

Repository Layout

• bnb/: branch-and-bound search, node batching, queue management, ADMM adapter, incumbent filtering, and debug logging.
• primal_dual_ilqr/: low-level structured trajectory QP solvers. The active backend is admm_tvlqr.py; fast_sls.py contains the migrated SLS backend.
• reach_avoid_bnb.py: single-target reach-avoid model, SCP loop, MPC loop, plotting helpers, and conversion into BNB problem data.
• ordered_reach_avoid_bnb.py: ordered multi-target reach-avoid variant.
• plot_bnb_reach_avoid.py: main CLI for the reach-avoid MPC/SCP examples.
• plot_ordered_reach_avoid.py: CLI for the ordered reach-avoid example.
• tests/: regression tests for ADMM wrapping, BNB search behavior, warm starts, infeasibility handling, SCP/MPC logic, plotting labels, and repository hygiene.
• example_outputs/reach_avoid_examples/: curated example plots and JSONL traces.

Environment

Run Python through the bnb conda environment:

conda run -n bnb python -V
conda run -n bnb python -m pytest tests -q

Use commands from the repository root:

cd /home/tailinfan/primal_dual_ilqr_stl_solver

Main Example

The primary reach-avoid MPC example is:

conda run -n bnb python plot_bnb_reach_avoid.py \
  --output example_outputs/reach_avoid_examples/reach_avoid_mpc4.png \
  --Np 2 --Ns 2 \
  --T 20 \
  --dt 1.0 \
  --admm-solver-mode adaptive \
  --rho-update-frequency 10 \
  --scp-max-iterations 4 \
  --mpc-steps 4

See USAGE.md for the full file guide, test commands, and the exact parameters used for the three checked example plots and JSONL traces.