Run the CaP Fix Loop + Evolutionary Search experiment on a user-supplied list of low-performing benchmark tasks: iterative K=8 candidate search on development seeds 51–65, validation selection, and final held-out evaluation on seeds 1–50.
The English summary is being prepared. Showing the original description.
Run score Not tested
See how to installReview the instructions the agent will receive before you install.
Use this skill for intensive debugging of difficult tasks — Evolutionary Search's
multi-candidate search finds strategies the single-pass fix loop misses. It builds on the
iterative-debugging skill: each task's fix loop fix_code.py (when present) is the baseline
to beat, seeded verbatim as candidate_A. LIBERO-Pro runs use --init-mode seeded so
development seeds 51–65 and held-out seeds 1–50 are distinct scenes.
docs/libero-pro.md,
or docs/robosuite.md), including the pinned environment (for LIBERO:
scripts/bootstrap_libero.sh → .venv-libero) and SAM3 authentication.LIBERO_CONFIG_PATH="$PWD/.libero",
MUJOCO_GL=egl, PYOPENGL_PLATFORM=egl, EGL_PLATFORM=device,
TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1).scripts/supervise_services.sh) and a persistent coordinator
session.fix_code.py (and a baseline held-out rate)
for the target tasks when available.docs/api-reference.md, docs/run-artifacts.md, and
main-agent-prompt.md.(benchmark, suite, task_id) cases) and follow
main-agent-prompt.md as the coordinator..claude/skills/iterative-debugging/scripts/run_validation.py before returning.| File | Purpose |
|---|---|
| main-agent-prompt.md | Coordinator runbook for task dispatch and result collection |
| subagent-prompt.md | Per-task Evolutionary Search iteration prompt |
| clean-task-slate.md | Reset checklist before reruns |
| skills/evosearch-iteration.md | Candidate/eval/keyframe iteration mechanics |
| scripts/evosearch_eval.py | Candidate × seed matrix evaluator |
| scripts/analyze_evosearch_traces.py | Post-eval per-candidate signal analysis |
docs/api-reference.md | Public CaP program API reference |
| ../iterative-debugging/skills/ | Shared robot skills (localize/grasp/transport/manipulation) |
No reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Localize objects and compose low-level CAP MCP actions using observations, Cartesian or synchronized motion, joint targets, and single-arm gripper control.
The English summary is being prepared. Showing the original description.
Run the baseline-free CaP iterative debugging (fix loop) pipeline on a user-supplied list of benchmark tasks: inspect one initial observed scene per task, generate task-level code, debug failures on development seeds using recorded run artifacts, validate on the held-out partition, grade the skill library against what actually worked, and promote only those patterns a verifier could reproduce the fix from.
The English summary is being prepared. Showing the original description.
Block-by-block fix loop on BEHAVIOR-1K (R1 Pro, OmniGibson on Isaac Sim). The policy is grown one code block at a time against what the robot actually did, and what carries across seeds is a skill library, not a program. Use for the two R1 Pro pickup tasks; Stage 1 learns on seeds 26-35, Stage 2 evaluates on 1-25 with a frozen library.
The English summary is being prepared. Showing the original description.
Perform task-specific real YAM manipulation with native CAP tools. Use for solving or turning a physical Rubik's Cube, bottle-cap removal and pouring, pill insertion and bottle closure, or inverted cup nesting.
The English summary is being prepared. Showing the original description.