onf/scripts/run_long_geo_full.sh

#!/usr/bin/env bash
# Full 7-axis `long` sweep under the geometric cleanliness rule, to sit beside the trained-field
# sweep already on disk (results/plus_libero_10/*/gr00t_{entry,sentinel}_long_*).
#
#   bash scripts/run_long_geo_full.sh [GPU] [PORT]
#
# GR_GRAPH_DIR is exported in THIS shell, not passed per-cell: run_gr00t resolves the graph dir once
# in the parent process, so a value that only appears later (e.g. a ladder cell's env block) is bound
# too late and the run silently falls back to outputs/long/latest -- which is what happened to the
# first StableVLA attempt.
#
# Robot_Initial_States is worth watching separately here: it runs the t=0 `entry` recipe, where the
# query window is 8 identical rows, so the cleanliness weight cancels out of the pooling and the
# entry aggregation discards the softmax entirely. That axis should come out ~unchanged; if it moves
# more than run-to-run noise, something other than the cleanliness rule differs between the arms.
set -uo pipefail

REPO="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
PY="${ONF_PY:-/home/quang/miniconda3/envs/stablevla/bin/python}"
GPU="${1:-2}"
PORT="${2:-10099}"
ART="$REPO/outputs/long/geo_v1/artifacts"

export OMP_NUM_THREADS=4 PYTHONPATH="$REPO/src"
export ONF_CLEANLINESS=geo GR_GRAPH_DIR="$ART"
cd "$REPO"; mkdir -p logs/_drivers

for f in g_nodes.npz g_edges.npz g_head.npz g_track.npz; do
    [ -f "$ART/$f" ] || { echo "[$(date -u +%FT%TZ)] !!! missing $ART/$f"; exit 1; }
done

echo "[$(date -u +%FT%TZ)] === long: 7 axes, geometric cleanliness, cuda:$GPU port $PORT"
echo "[$(date -u +%FT%TZ)]     graph dir: $ART"
$PY evals/gr00t/run_gr00t.py --suites long --gpu "$GPU" --port "$PORT" --num-clients 6 \
    --variant geo_full >>logs/_drivers/gr00t_long_geo_full.log 2>&1
echo "[$(date -u +%FT%TZ)] === exit $?"