strict640x480-v2/code/wait_and_evaluate_all5.sh

#!/usr/bin/env bash
# Queue a formal 100-seed All-5 evaluation without racing the corresponding
# training process.  Each watcher is pinned to the GPU vacated by its method.
set -euo pipefail

method="${1:?method: dino|stereo|moe|local_arca}"
gpu="${2:?physical GPU index}"
skip_tasks=",${3:-},"
eval_workers="${EVAL_WORKERS:-1}"
root=/workspace/RoboFactory/runs/strict640x480_v2
py=/workspace/venvs/robofactory-act/bin/python3.10

case "$method" in
  dino)
    checkpoint="$root/results/frozen_dinov3_act_all5_80k/checkpoint_080000.pt"
    evaluator=/workspace/act_liftbarrier/evaluate_shared_act.py
    result_dir="$root/results/frozen_dinov3_act_all5_80k/formal_heldout_100"
    ;;
  stereo)
    checkpoint="$root/results/stereo_cross_relbias_all5_80k/checkpoint_080000.pt"
    evaluator=/workspace/act_liftbarrier/evaluate_stereo_act.py
    result_dir="$root/results/stereo_cross_relbias_all5_80k/formal_heldout_100"
    ;;
  moe)
    checkpoint="$root/results/stereo_ffn_moe_all5_80k/checkpoint_080000.pt"
    evaluator=/workspace/act_liftbarrier/evaluate_stereo_act.py
    result_dir="$root/results/stereo_ffn_moe_all5_80k/formal_heldout_100"
    ;;
  local_arca)
    checkpoint="$root/results/local_arca_all5_80k/checkpoint_080000.pt"
    evaluator=/workspace/act_liftbarrier/evaluate_stereo_act.py
    result_dir="$root/results/local_arca_all5_80k/formal_heldout_100"
    ;;
  *) echo "unknown method: $method" >&2; exit 2 ;;
esac

while [[ ! -s "$checkpoint" ]]; do
  sleep 60
done

mkdir -p "$result_dir"
for task in lift_barrier camera_alignment three_robots_stack_cube long_pipeline_delivery take_photo; do
  case "$skip_tasks" in *",$task,"*) echo "SKIP_ASSIGNED_TASK $method $task"; continue;; esac
  result="$result_dir/eval_${task}.json"
  # A parallel worker may have already completed this exact frozen protocol on
  # another freed GPU.  Reuse only a complete 100-episode JSON; never treat a
  # partial file as valid.
  if [[ -s "$result" ]] && "$py" - "$result" <<'PY'
import json, sys
r = json.load(open(sys.argv[1], encoding="utf-8"))
raise SystemExit(0 if r.get("episodes") == 100 and len(r.get("rows", [])) == 100 else 1)
PY
  then
    echo "REUSE_FORMAL_EVAL $method $task"
    continue
  fi
  # A single worker owns the freed GPU, avoiding concurrent simulator processes
  # silently contending for one policy replica.  The seed file is frozen and
  # verified by the evaluator before any rollout starts.
  CUDA_VISIBLE_DEVICES="$gpu" "$py" -u "$evaluator" \
    --checkpoint "$checkpoint" --task "$task" \
    --seed-file "$root/heldout_seeds/$task.json" --episodes 100 \
    --workers "$eval_workers" --devices 0 --max-steps 1500 \
    --output "$result"
done

printf 'FORMAL_ALL5_EVALUATION_COMPLETE %s\n' "$method"