train_robosuite_phase1.sh

#!/bin/bash
# Phase 1: Train Lift, then Stack sequentially (demo subgoals only; vgen disabled).
# Each task runs to completion (2M steps) unless you Ctrl+C.

set -e
cd "$(dirname "$0")"

SEED="${SEED:-0}"
DEVICE="${DEVICE:-cuda:0}"
PRETRAIN_ROOT="${PRETRAIN_ROOT:-/home/lei/Documents/tong/irl4idm/multi-task-tcc-robosuite/experiments/pretrain_runs}"
LIFT_ENCODER="${LIFT_ENCODER:-${PRETRAIN_ROOT}/dataset=mimicgen_algo=xirl_task=lift}"
STACK_ENCODER="${STACK_ENCODER:-${PRETRAIN_ROOT}/dataset=mimicgen_algo=xirl_task=stack}"

run_task() {
  local env_name="$1"
  local exp_name="$2"
  local encoder_path="$3"
  local resume_flag=()
  local ckpt_dir="rl_runs/${exp_name}/${SEED}/checkpoints"

  if compgen -G "${ckpt_dir}/*.ckpt" > /dev/null; then
    resume_flag=(--resume)
    echo "Found checkpoint in ${ckpt_dir}; resuming."
  fi

  echo ""
  echo "=========================================="
  echo " Phase 1: ${env_name}"
  echo " Experiment: rl_runs/${exp_name}/${SEED}"
  echo " Encoder: ${encoder_path}"
  echo " TensorBoard: tensorboard --logdir rl_runs/${exp_name}/${SEED}/tb"
  echo "=========================================="
  echo ""

  python train_policy.py \
    --experiment_name="${exp_name}" \
    --env_name="${env_name}" \
    --config="configs/robosuite/rl/env_reward.py:${env_name}" \
    --config.reward_wrapper.pretrained_path="${encoder_path}" \
    --config.reward_wrapper.type=distance_to_goal \
    --config.threshold_for_vgen=1.1 \
    --seed="${SEED}" \
    --device="${DEVICE}" \
    "${resume_flag[@]}"
}

run_task Lift \
  "${LIFT_EXP_NAME:-lift_rl_demo_only}" \
  "${LIFT_ENCODER}"

run_task Stack \
  "${STACK_EXP_NAME:-stack_rl_demo_only}" \
  "${STACK_ENCODER}"

echo ""
echo "Done: Lift and Stack phase 1 complete."
echo "After your video models are ready, run: ./train_robosuite_phase2_resume.sh"