feat: LIBERO benchmark adapter + BDDL parser (#110)

[yinsong1986]

Summary

Implements #110 as a follow-up to #129 (the BenchmarkProtocol framework from #107). Ships LiberoAdapter so the ~130 LIBERO tasks become runnable on the MuJoCo backend through the standard evaluate_benchmark surface:

sim.evaluate_benchmark(
    benchmark_name="libero-spatial-pick_up_the_red_cube",
    policy_provider="lerobot_local",
    n_episodes=50,
    seed=42,
)

Stacked on #129. #129 must merge first. The diff below will include #129's commits until then; once #129 merges this PR auto-rebases onto main.

What's in scope (from #110)

1. BDDL parser (strands_robots/benchmarks/libero/bddl_parser.py) — pure-Python tokenizer + s-expression parser + typed AST + goal compiler. Handles define, :domain, :objects, :init, :goal, :language, boolean combinators (and / or / not), and the closed predicate vocabulary LIBERO uses: on, near, inside, open, closed, grasped, upright. Unknown predicates rejected explicitly with the valid list. No eval(), no exec() — safe for untrusted input.
2. Scene loading — LiberoAdapter.on_episode_start loads the declared scene MJCF (if provided) before the base compatibility check, so the check sees the scene's Panda robot rather than reporting "sim empty → load default".
3. LiberoAdapter class (strands_robots/benchmarks/libero/adapter.py) — Panda-only BenchmarkProtocol subclass. Compiles BDDL goal → single (sim) -> bool callable. Sparse on_step (LIBERO has no dense reward). Applies per-episode RNG-seeded ±jitter to init-subject bodies.
4. Registry bootstrap — load_libero_suite(suite_name) bulk-registers every BDDL task in a suite under libero-<suite>-<task> keys. Accepts bddl_dir= override so the adapter is fully usable without the libero pip package — you can ship your own BDDL files. LIBERO is only imported when auto-discovering task files from the installed package.
5. Packaging — new optional extra strands-robots[benchmark-libero] with libero>=0.1.0,<1.0.0. Not added to [all] (LIBERO is a heavy opt-in dep). Core tests pass without it installed.
6. Tests — 91 new tests (see below) covering parser, adapter lifecycle, suite loader, and end-to-end MuJoCo dispatch.

New generic predicates

BDDL requires 4 predicates not yet in the core library; they're generic enough to live there rather than being LIBERO-specific:

Predicate	Signature
body_on(body_a, body_b, z_offset=0.02, xy_tol=0.15)	A is above B and horizontally close
body_inside(body, container, xy_tol=0.15, z_tol=0.15)	A is within B's AABB
body_upright(body, tol=0.15)	body's local +Z axis within tol of world +Z
grasped(body, gripper_prefix)	body in contact with any geom matching gripper_prefix*

All use only SimEngine.get_body_state / get_contacts so they remain backend-agnostic.

Architecture notes

• Panda-only by design. LIBERO's scene MJCFs <include> Panda geometry and BDDL predicates reference Panda gripper body names (robot0_gripper_*). Retargeting to SO-100 et al. would require rewriting every BDDL predicate against different body names — out of scope.
• BDDL parser never runs Python. Closed token set + closed predicate vocabulary. LLM-authored or agent-authored BDDL files are safe to parse.
• Malformed BDDL is logged, not fatal. load_libero_suite keeps loading the rest of the suite if one file is bad — so one upstream drift doesn't brick a whole suite.

Tests (91 new, all passing)

• tests/benchmarks/libero/test_bddl_parser.py (38) — tokenizer / parser / vocabulary / combinators / round-trip on one example per predicate family.
• tests/benchmarks/libero/test_libero_adapter.py (19) — construction, Panda-only compat, is_success cases, scene-load ordering, init jitter determinism, PolicyRunner / evaluate_benchmark integration, unknown-task structured error.
• tests/benchmarks/libero/test_libero_suite.py (15) — suite name normalisation, bulk registration, scene resolution, malformed-BDDL skip, forwarding of max_steps / init_jitter, error cases.
• tests/benchmarks/libero/test_libero_e2e.py (2) — end-to-end MuJoCo dispatch path.
• tests/simulation/test_benchmark_predicates.py (+17) — new predicate tests (body_on / body_inside / body_upright / grasped).

Verification

hatch run format                  # ruff check --fix + format
hatch run lint                    # all clean
hatch run pytest tests/           # 1438 passed, 31 skipped (env-dependent GL skips)

Without LIBERO installed:

python -c "
import strands_robots
from strands_robots.benchmarks.libero import LiberoAdapter, parse_bddl
a = LiberoAdapter.from_text('(define (problem t) (:goal (grasped cube)))')
print(a.problem.name, a.supported_robots)  # → t ['panda']
"
# → ok
# load_libero_suite without bddl_dir= raises a clean ImportError naming [benchmark-libero].

Example: register + evaluate a LIBERO task

from strands_robots.simulation import Simulation
from strands_robots.benchmarks.libero import load_libero_suite

sim = Simulation(tool_name="libero_sim", mesh=False)
sim.create_world()

# Bulk-register every task in libero_spatial (requires libero pip pkg OR bddl_dir=)
load_libero_suite("libero_spatial", scene_dir="/path/to/scenes")

result = sim.evaluate_benchmark(
    benchmark_name="libero-spatial-pick_up_the_red_cube",
    policy_provider="lerobot_local",
    policy_config={"pretrained_name_or_path": "lerobot/smolvla"},
    n_episodes=20,
    seed=42,
)

Out of scope (per #110)

• Dense-reward variants of LIBERO tasks (LIBERO is sparse-success by design).
• Language-conditioning evaluation tooling (adapter.instruction surfaces the BDDL :language string; the policy consumes it as-is).
• Hosting LIBERO's MJCF assets (depend on upstream).

Closes #110.

---

Reviewers from #129 — would appreciate your eyes on this one too: @sundargthb @awsarron @cagataycali

[yinsong1986]

Requesting review from @sundargthb, @awsarron, @cagataycali 🙏