diff --git a/.DS_Store b/.DS_Store
new file mode 100644
index 0000000..b46e77c
Binary files /dev/null and b/.DS_Store differ
diff --git a/.libero/config.yaml b/.libero/config.yaml
new file mode 100644
index 0000000..4fa7ba6
--- /dev/null
+++ b/.libero/config.yaml
@@ -0,0 +1,5 @@
+assets: /home/xsuper/dsrl_pi0/LIBERO/libero/libero/assets
+bddl_files: /home/xsuper/dsrl_pi0/LIBERO/libero/libero/bddl_files
+benchmark_root: /home/xsuper/dsrl_pi0/LIBERO/libero/libero
+datasets: /home/xsuper/dsrl_pi0/LIBERO/libero/datasets
+init_states: /home/xsuper/dsrl_pi0/LIBERO/libero/libero/init_files
diff --git a/.vscode/settings.json b/.vscode/settings.json
new file mode 100644
index 0000000..4b5a294
--- /dev/null
+++ b/.vscode/settings.json
@@ -0,0 +1,4 @@
+{
+    "python-envs.defaultEnvManager": "ms-python.python:conda",
+    "python-envs.defaultPackageManager": "ms-python.python:conda"
+}
\ No newline at end of file
diff --git a/examples/scripts/run_libero.sh b/examples/scripts/run_libero.sh
index 7e0f2b5..3bb3bd2 100644
--- a/examples/scripts/run_libero.sh
+++ b/examples/scripts/run_libero.sh
@@ -1,16 +1,21 @@
 #!/bin/bash
 proj_name=DSRL_pi0_Libero
 device_id=0
-
+#fff
 export DISPLAY=:0
 export MUJOCO_GL=egl
 export PYOPENGL_PLATFORM=egl  
 export MUJOCO_EGL_DEVICE_ID=$device_id
-
 export OPENPI_DATA_HOME=./openpi
 export EXP=./logs/$proj_name; 
 export CUDA_VISIBLE_DEVICES=$device_id
 export XLA_PYTHON_CLIENT_PREALLOCATE=false
+export PIP_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple
+export PYTHONPATH=$PWD:$PWD/LIBERO:$PWD/openpi/src:$PYTHONPATH
+export LIBERO_CONFIG_PATH=$PWD/.libero
+export NUMBA_DISABLE_JIT=1
+export MPLCONFIGDIR=$PWD/.cache/matplotlib
+mkdir -p "$MPLCONFIGDIR"
 
 pip install mujoco==3.3.1
 
@@ -31,4 +36,4 @@ python3 examples/launch_train_sim.py \
 --resize_image 64 \
 --action_magnitude 1.0 \
 --query_freq 20 \
---hidden_dims 128 \
\ No newline at end of file
+--hidden_dims 128 \
diff --git a/examples/train_sim.py b/examples/train_sim.py
index a9d99c9..a992137 100755
--- a/examples/train_sim.py
+++ b/examples/train_sim.py
@@ -1,9 +1,11 @@
 #! /usr/bin/env python
 import os
-# Tell XLA to use Triton GEMM, this improves steps/sec by ~30% on some GPUs from https://github.com/huggingface/gym-aloha/tree/main?tab=readme-ov-file#-gpu-rendering-egl
-xla_flags = os.environ.get('XLA_FLAGS', '')
-xla_flags += ' --xla_gpu_triton_gemm_any=True'
-os.environ['XLA_FLAGS'] = xla_flags
+# Triton GEMM can improve throughput on some GPUs, but older JAX/XLA builds may
+# crash while compiling half-precision kernels on newer architectures.
+if os.environ.get('DSRL_ENABLE_TRITON_GEMM', '0') == '1':
+    xla_flags = os.environ.get('XLA_FLAGS', '')
+    xla_flags += ' --xla_gpu_triton_gemm_any=True'
+    os.environ['XLA_FLAGS'] = xla_flags
 
 import pathlib, copy
 
@@ -161,4 +163,4 @@ def main(variant):
     replay_buffer = online_replay_buffer
     replay_buffer.seed(variant.seed)
     trajwise_alternating_training_loop(variant, agent, env, eval_env, online_replay_buffer, replay_buffer, wandb_logger, shard_fn=shard_fn, agent_dp=agent_dp)
- 
\ No newline at end of file
+