code/run_ood_groot_inference.sh

#!/usr/bin/env bash
set -euo pipefail
# ─────────────────────────────────────────────────────────────────────────────
# run_ood_groot_inference.sh   (GR00T N1.7 version)
#
# Mirrors genie_envisioner/run_ood_experiment_inference.sh EXACTLY (same job
# generation / pair+seed sampling / results format) but the policy is a
# fine-tuned GR00T N1.7 checkpoint served over zmq by run_gr00t_server.py.
#
# The GR00T inference server must already be running and reachable at
# ${HOST}:${PORT} (run_all_groot.sh starts/stops one per checkpoint).
#
# Usage:
#   bash run_ood_groot_inference.sh <experiment> [seed] [total_episodes] [results_txt_path]
#
# experiment:
#   verb_color | verb_object | color_object | verb_size | verb_spatial |
#   size_object | color_size | color_spatial | spatial_size | spatial_object
#
# seed (default: 42)            RNG seed for pair / third-factor / episode seeds
# total_episodes (default: 200) #(i,j) pairs sampled; total_runs = 2 × this
# results_txt_path              summary log (default: auto-timestamped)
#
# Env vars:
#   HOST 127.0.0.1   PORT 5555            GR00T server address
#   SIM_BACKEND gpu                       ManiSkill physics+render backend
#   MAX_EPISODE_STEPS 300   REPLAN_STEPS 5
#   SEED_BASE 0
#   EXPERIMENT_ROOT data/conflict_groot/experiments   (videos saved here)
#   GROOT_MAIN  path to groot_main.py
#   MS_PY       python interpreter of the ManiSkill venv
# ─────────────────────────────────────────────────────────────────────────────

EXPERIMENT="${1:-}"
SEED="${2:-42}"
TOTAL_EPISODES_TARGET="${3:-200}"
RESULTS_TXT_PATH="${4:-}"
THIRD_SEED="${SEED}"

if [[ -z "${EXPERIMENT}" ]]; then
  echo "Usage: $0 <experiment> [seed] [total_episodes] [results_txt_path]"
  exit 1
fi
case "${EXPERIMENT}" in
  verb_color|verb_object|color_object|verb_size|verb_spatial|\
  size_object|color_size|color_spatial|spatial_size|spatial_object) ;;
  *) echo "Unsupported experiment: ${EXPERIMENT}"; exit 1 ;;
esac
if ! [[ "${TOTAL_EPISODES_TARGET}" =~ ^[0-9]+$ ]] || [[ "${TOTAL_EPISODES_TARGET}" -le 0 ]]; then
  echo "total_episodes must be a positive integer"; exit 1
fi

HOST="${HOST:-127.0.0.1}"
PORT="${PORT:-5555}"
SIM_BACKEND="${SIM_BACKEND:-gpu}"
MAX_EPISODE_STEPS="${MAX_EPISODE_STEPS:-300}"
REPLAN_STEPS="${REPLAN_STEPS:-5}"
SEED_BASE="${SEED_BASE:-0}"
EXPERIMENT_ROOT="${EXPERIMENT_ROOT:-data/conflict_groot/experiments}"
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
GROOT_MAIN="${GROOT_MAIN:-${SCRIPT_DIR}/groot_main.py}"
MS_PY="${MS_PY:-/workspace/groot_eval/.venv_ms/bin/python}"

if [[ -z "${RESULTS_TXT_PATH}" ]]; then
  ts="$(date +%Y%m%d_%H%M%S)"
  RESULTS_TXT_PATH="${EXPERIMENT_ROOT}/${EXPERIMENT}_ood_${ts}.txt"
fi
mkdir -p "$(dirname "${RESULTS_TXT_PATH}")"

# ── Generate jobs JSON — VERBATIM logic from genie run_ood_experiment_inference.sh ──
JOBS_JSON="$(mktemp --suffix=.json)"
python3 - "${EXPERIMENT}" "${SEED}" "${TOTAL_EPISODES_TARGET}" "${SEED_BASE}" "${THIRD_SEED}" <<'PY' > "${JOBS_JSON}"
import random, sys, json, math
experiment  = sys.argv[1]
seed        = int(sys.argv[2])
n_episodes  = int(sys.argv[3])
seed_base   = int(sys.argv[4])
third_seed  = int(sys.argv[5])
rng = random.Random(seed)
def _size_swap(n_size):
    p = []
    for a, b in ((0,1),(2,3),(4,5)):
        if a < n_size and b < n_size:
            p += [(a,b),(b,a)]
    return p
_SIZE_EXPS    = {"verb_size", "size_object", "color_size"}
_SPATIAL_EXPS = {"verb_spatial", "color_spatial", "spatial_size", "spatial_object"}
if experiment in _SIZE_EXPS:
    all_pairs = _size_swap(6)
elif experiment == "spatial_size":
    all_pairs = _size_swap(5)
elif experiment in _SPATIAL_EXPS:
    n = 5
    all_pairs = [(i,j) for i in range(n) for j in range(n) if i != j]
else:
    n = 6
    all_pairs = [(i,j) for i in range(n) for j in range(n) if i != j]
_run_types = {
    "verb_color":("verb","color"), "verb_object":("verb","shape"),
    "verb_size":("verb","size"),   "verb_spatial":("verb","spatial"),
    "color_object":("color","shape"), "size_object":("size","shape"),
    "color_size":("color","size"), "color_spatial":("color","spatial"),
    "spatial_size":("spatial","size"), "spatial_object":("spatial","shape"),
}
first, second = _run_types[experiment]
raw_jobs = []
for ep_idx in range(n_episodes):
    i, j = rng.choice(all_pairs)
    raw_jobs.append((i, j, first, ep_idx))
    raw_jobs.append((i, j, second, ep_idx))
total = len(raw_jobs)
num_episodes = math.ceil(n_episodes / total)
jobs = []
for k, (i, j, run_type, ep_idx) in enumerate(raw_jobs):
    idx = k + 1
    run_name = f"ood_{idx:03d}_{experiment}_{i}_{j}_{run_type}"
    if experiment == "verb_object":
        run_seed = seed_base + ep_idx
        ep_third_seed = ep_idx
    else:
        run_seed = seed_base + idx
        ep_third_seed = third_seed
    jobs.append({
        "index": idx, "pair_i": i, "pair_j": j, "run_type": run_type,
        "seed": run_seed, "third_seed": ep_third_seed,
        "num_episodes": num_episodes, "experiment_name": run_name,
    })
print(json.dumps(jobs))
PY

read -r TOTAL NUM_EPISODES < <(python3 - "${JOBS_JSON}" <<'PY'
import json, sys
jobs = json.loads(open(sys.argv[1]).read())
print(len(jobs), jobs[0]["num_episodes"] if jobs else 1)
PY
)
echo "Requested ${TOTAL_EPISODES_TARGET} total episodes across ${TOTAL} runs → ${NUM_EPISODES} eps/run (actual: $((TOTAL * NUM_EPISODES)))"

{
  echo "# OOD pairwise inference summary (GR00T N1.7)"
  echo "experiment=${EXPERIMENT}"
  echo "seed=${SEED}"
  echo "total_episodes_target=${TOTAL_EPISODES_TARGET}"
  echo "num_episodes_per_run=${NUM_EPISODES}"
  echo "total_runs=${TOTAL}"
  echo "total_episodes_actual=$((TOTAL * NUM_EPISODES))"
  echo "third_seed=${THIRD_SEED}"
  echo "groot_server=${HOST}:${PORT}"
  echo "sim_backend=${SIM_BACKEND}"
  echo "max_episode_steps=${MAX_EPISODE_STEPS}"
  echo "replan_steps=${REPLAN_STEPS}"
  echo "seed_base=${SEED_BASE}"
  echo
  echo "index pair_i pair_j run_type success run_name"
} > "${RESULTS_TXT_PATH}"

# ManiSkill C-extensions (fast_kinematics/mplib) link libtorch.so — make the
# ms-venv torch libs discoverable regardless of caller.
_MS_TORCH_LIB="$("${MS_PY}" -c 'import torch,os;print(os.path.join(os.path.dirname(torch.__file__),"lib"))' 2>/dev/null || true)"
export LD_LIBRARY_PATH="${_MS_TORCH_LIB}:${LD_LIBRARY_PATH:-}"
export MANISKILL_CONFLICT_ROOT="${MANISKILL_CONFLICT_ROOT:-/workspace/groot_eval/genie_repo/maniskill_conflict}"

"${MS_PY}" "${GROOT_MAIN}" \
  --experiment          "${EXPERIMENT}" \
  --host                "${HOST}" \
  --port                "${PORT}" \
  --replan-steps        "${REPLAN_STEPS}" \
  --max-episode-steps   "${MAX_EPISODE_STEPS}" \
  --sim-backend         "${SIM_BACKEND}" \
  --experiment-root     "${EXPERIMENT_ROOT}" \
  --batch-jobs-file     "${JOBS_JSON}" \
  --batch-results-txt   "${RESULTS_TXT_PATH}"
py_status=$?
rm -f "${JOBS_JSON}"
if [[ "${py_status}" -ne 0 ]]; then
  echo "Batch eval failed (exit ${py_status}); partial results in ${RESULTS_TXT_PATH}"
  exit "${py_status}"
fi
echo "Saved summary to ${RESULTS_TXT_PATH}"