1source switch_env <env_name>
2# source switch_env openvla
3# source switch_env openvla-oft
4# source switch_env openpi

1git clone https://github.com/RLinf/RLinf
2# 为提高国内下载速度，可以使用：
3# git clone https://ghfast.top/github.com/RLinf/RLinf.git

1pip install huggingface-hub
2# 此处可以调整自己的下载路径
3# 例如 pi05 模型
4hf download RLinf/RLinf-Pi05-LIBERO-SFT --local-dir ./data_asserts/
5# pi0 模型
6# hf download RLinf/RLinf-Pi0-LIBERO-Spatial-Object-Goal-SFT --local-dir ./data_asserts/

1cluster:
2   num_nodes: 1
3   component_placement:
4      env: 0-3
5      rollout: 4-7
6      actor: 0-7
7
8rollout:
9   pipeline_stage_num: 2

1cluster:
2   num_nodes: 1
3   component_placement:
4      env,rollout,actor: all

1cluster:
2   num_nodes: 1
3   component_placement:
4      env: 0-1
5      rollout: 2-5
6      actor: 6-7

1openpi:
2  noise_level: 0.5
3  action_chunk: ${actor.model.num_action_chunks}
4  num_steps: ${actor.model.num_steps}
5  train_expert_only: True
6  action_env_dim: ${actor.model.action_dim}
7  noise_method: "flow_sde"
8  add_value_head: False
9  pi05: False
10  value_after_vlm: False

1model:
2  is_lora: True
3  lora_rank: 8
4  gradient_checkpointing: False

1# 使用模型下载中使用地址的绝对路径
2model_dir: "./data_asserts/"

1# 通过调整 CHOSEN_CONFIG 调用不同训练任务
2# bash examples/embodiment/run_embodiment.sh CHOSEN_CONFIG
3bash examples/embodiment/run_embodiment.sh libero_10_ppo_openpi_pi05

1# 可将 ./logs 改为日志地址
2tensorboard --logdir ./logs --port 6006

1.
2├── libero_10_ppo_openpi_pi05
3│   └── checkpoints
4│       └── global_step_40
5│           └── actor
6│               ├── dcp_checkpoint
7│               └── model_state_dict
8├── tensorboard
9└── video
10    └── train
11        ├── seed_0
12        ├── seed_1
13        ├── seed_2
14        ├── seed_3
15        ├── seed_4
16        ├── seed_5
17        ├── seed_6
18        └── seed_7

1#! /bin/bash
2
3export EMBODIED_PATH="$( cd "$(dirname "${BASH_SOURCE[0]}" )" && pwd )"
4export REPO_PATH=$(dirname "$(dirname "$EMBODIED_PATH")")
5export SRC_FILE="${EMBODIED_PATH}/eval_embodied_agent.py"
6
7export MUJOCO_GL="osmesa"
8export PYOPENGL_PLATFORM="osmesa"
9export PYTHONPATH=${REPO_PATH}:$PYTHONPATH
10
11# LIBERO 仓库路径
12export LIBERO_PATH="/opt/LIBERO"
13export PYTHONPATH=${LIBERO_PATH}:$PYTHONPATH
14
15export CUDA_LAUNCH_BLOCKING=1
16export HYDRA_FULL_ERROR=1
17
18CONFIG_NAME=${1:-libero_goal_grpo_openvlaoft.eval}
19
20LOG_DIR="${REPO_PATH}/logs/$(date +'%Y%m%d-%H:%M:%S')"
21MEGA_LOG_FILE="${LOG_DIR}/eval_embodiment.log"
22mkdir -p "${LOG_DIR}"
23
24CMD="python ${SRC_FILE} --config-path ${EMBODIED_PATH}/config/ \
25     --config-name ${CONFIG_NAME} \
26     runner.logger.log_path=${LOG_DIR}"
27
28echo ${CMD}
29${CMD} 2>&1 | tee "${MEGA_LOG_FILE}"

1runner:
2  eval_policy_path: "/path/to/rl_ckpt.pt"    # 可选：.pt文件路径或直接填None。如果为None，将使用rollout.model.model_path中的检查点。
3  only_eval: True    # 配置为 true 这样就不会再 eval 中间再跑 training 了
4algorithm:
5  eval_rollout_epoch: 1
6rollout:
7  model:
8    model_path: "/path/to/sft_base_model/"
9actor:
10  model:
11    model_path: "/path/to/sft_base_model/"
12  tokenizer:
13    tokenizer_model: "/path/to/sft_base_model/"

1bash examples/embodiment/eval_embodiment.sh libero_10_ppo_openpi_pi05

1# export MUJOCO_GL="osmesa"
2# export PYOPENGL_PLATFORM="osmesa"
3
4export MUJOCO_GL="egl"
5export PYOPENGL_PLATFORM="egl"

1.
2├── tensorboard
3└── video
4    └── eval
5        ├── seed_0
6        ├── seed_1
7        ├── seed_2
8        ├── seed_3
9        ├── seed_4
10        ├── seed_5
11        ├── seed_6
12        └── seed_7

1eval_metrics={
2    'eval/success_once': array(0.9173387, dtype=float32), 
3    'eval/return': array(0.7520161, dtype=float32), 
4    'eval/episode_len': array(480., dtype=float32), 
5    'eval/reward': array(0.0015667, dtype=float32), 
6    'eval/success_at_end': array(0.7520161, dtype=float32)
7   }

1reward: 0.0                        # 这一步环境给的即时奖励
2terminated: false                  # 当前task是否完成
3task: pick up the book             # 子任务描述
4and place it in the
5back compartment of
6the caddy