# LIBERO-Para (/zh/docs/evaluation/benchmark-hub/libero-para)



## 简介 [#简介]

LIBERO-Paraphrase 测试 VLA 策略在任务语言 paraphrase 下的泛化能力。

## 指标 [#指标]

LIBERO-Para external benchmark task manifest for paraphrased language-conditioned robot manipulation evaluation.

### Task YAML [#task-yaml]

`worldfoundry/data/benchmarks/tasks/external/libero-para.yaml`

### Primary [#primary]

* `success_rate`

| Metric                            | 关注点                              |
| --------------------------------- | -------------------------------- |
| `success_rate`                    | episode / task 是否成功完成的比例或成功率。    |
| `paraphrase_success_rate`         | episode / task 是否成功完成的比例或成功率。    |
| `language_generalization_success` | episode / task 是否成功完成的比例或成功率。    |
| `task_success`                    | episode / task 是否成功完成的比例或成功率。    |
| `action_accuracy`                 | 预测动作、状态或结果相对 benchmark 成功标准的正确性。 |

## 环境准备 [#环境准备]

### 运行环境 [#运行环境]

* **Conda 环境：** `giga-world-policy-py311`。
* **单独配置：** 默认复用 WorldFoundry 统一环境（`needs_new_env: false`），通常不需要额外 benchmark-only conda env。
* LIBERO-Para is tracked as an embodied benchmark, not a model runner; WorldFoundry normalizes official-shaped result exports only.
* Full official status requires the cau-hai-lab/LIBERO-Para runtime, HAI-Lab/LIBERO-Para data, policy checkpoint, and paraphrase split mapping.

### 测评数据 [#测评数据]

* 待测模型输出目录：设置 `WORLDFOUNDRY_LIBERO_PARA_RESULTS_PATH` 指向生成视频或 rollout artifact 根目录。
* `${WORLDFOUNDRY_HFD_DATASET_ROOT}/HAI-Lab__LIBERO-Para`
* `${WORLDFOUNDRY_CACHE_DIR}/repos/cau-hai-lab--LIBERO-Para`

### Checkpoint 与资产 [#checkpoint-与资产]

* Base-model / metric 依赖：`libero_para_source_assets`, `libero_para_dataset_assets`。
* Metric / evaluator checkpoint 默认路径与覆盖变量见 [local assets 指南](/zh/docs/guides/local-assets)。

### 关键环境变量 [#关键环境变量]

* `WORLDFOUNDRY_LIBERO_PARA_ROOT`
* `WORLDFOUNDRY_LIBERO_PARA_RESULTS_PATH`
* `WORLDFOUNDRY_LIBERO_PARA_DATASET_ROOT`

### 环境检查 [#环境检查]

```bash
PYTHONPATH=${WORLDFOUNDRY_CACHE_DIR}/repos/cau-hai-lab--LIBERO-Para:src ${WORLDFOUNDRY_CONDA_ENVS_ROOT}/worldfoundry-giga-world-policy-py311/bin/python -m worldfoundry.evaluation.tasks.execution.orchestration.runtime_preflight --profile libero-para --manifest worldfoundry/data/benchmarks/runtime_profiles/official/libero-para.yaml --output-dir tmp/worldfoundry_preflight/libero-para --json
```

## 运行评测 [#运行评测]

准备好该 benchmark 页面列出的资产和候选产物后，使用下面的公开运行入口。

### 设置候选产物目录 [#设置候选产物目录]

```bash
export WORLDFOUNDRY_GENERATED_ARTIFACT_DIR=/path/to/generated/artifacts
```

### 导入已有官方结果 [#导入已有官方结果]

```bash
worldfoundry-eval zoo benchmark-run \
  --benchmark-id libero-para \
  --mode normalizer \
  --official-results-path /path/to/official/results-or-report \
  --generated-artifact-dir "${WORLDFOUNDRY_GENERATED_ARTIFACT_DIR}" \
  --output-dir tmp/libero-para/normalizer \
  --json
```

### 从生成产物重新打分 [#从生成产物重新打分]

```bash
worldfoundry-eval zoo benchmark-run \
  --benchmark-id libero-para \
  --mode official-run \
  --generated-artifact-dir "${WORLDFOUNDRY_GENERATED_ARTIFACT_DIR}" \
  --output-dir tmp/libero-para/official-run \
  --json
```

### 直接调用仓内 runner [#直接调用仓内-runner]

```bash
worldfoundry-eval zoo benchmark-run --benchmark-id libero-para --mode official-validation --official-results-path <official_results.json> --output-dir <out> --json
```

## 依赖与产物 [#依赖与产物]

### Inputs [#inputs]

* Official LIBERO-Para repository checkout
* HAI-Lab/LIBERO-Para dataset root
* Policy checkpoint for selected upstream eval script
* Official result dump from metrics/analyze\_results.py or equivalent structured export

### Environment [#environment]

* `WORLDFOUNDRY_LIBERO_PARA_RESULTS_PATH`

### Outputs [#outputs]

* `scorecard.json`
* `benchmark_contract.json`
* `raw_metric_table.jsonl`
* `raw_results.jsonl`

[← 返回 Benchmark Hub](/zh/docs/evaluation/benchmark-hub)