#!/usr/bin/env bash
# Run DreamerV3-torch on AntMaze-Medium-Play with exploitation logging.
# Usage: bash run_antmaze.sh [logdir]
set -e

LOGDIR="${1:-./logdir/antmaze_medium_play}"
mkdir -p "$LOGDIR"

cd "$(dirname "$0")/dreamerv3-torch"

echo "=== Starting DreamerV3 on AntMaze-Medium-Play ==="
echo "    logdir: $LOGDIR"
echo "    GPU:    $(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>/dev/null || echo 'N/A')"
echo ""

# Key TensorBoard metrics to watch for model exploitation:
#   imag_reward_mean  — policy imaginary reward (in-model)
#   env_reward_mean   — actual environment reward (from replay buffer)
#   kl                — RSSM posterior/prior divergence
#   reward_loss       — world model reward head loss
#
# Exploitation signal: imag_reward_mean >> env_reward_mean and growing gap

conda run -n dreamer python dreamer.py \
  --configs antmaze \
  --task antmaze_medium-play \
  --logdir "$LOGDIR" \
  --seed 0

echo ""
echo "=== Done. Launch TensorBoard with: ==="
echo "    conda run -n dreamer tensorboard --logdir $LOGDIR"