diff --git a/.github/workflows/benchmark_tests.yml b/.github/workflows/benchmark_tests.yml
new file mode 100644
index 000000000..79d5614b2
--- /dev/null
+++ b/.github/workflows/benchmark_tests.yml
@@ -0,0 +1,312 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Integration tests: build an isolated Docker image per benchmark and run a
+# 1-episode smoke eval. Each benchmark gets its own image so incompatible
+# dependency trees (e.g. hf-libero vs metaworld==3.0.0) can never collide.
+#
+# To add a new benchmark:
+#   1. Add docker/Dockerfile.benchmark.<name>  (install only lerobot[<name>])
+#   2. Copy one of the jobs below and adjust the image name and eval command.
+name: Benchmark Integration Tests
+
+on:
+  # Run manually from the Actions tab
+  workflow_dispatch:
+
+  # Run every Monday at 02:00 UTC.
+  schedule:
+    - cron: "0 2 * * 1"
+
+  push:
+    branches:
+      - main
+    paths:
+      - "src/lerobot/envs/**"
+      - "src/lerobot/scripts/lerobot_eval.py"
+      - "docker/Dockerfile.benchmark.*"
+      - ".github/workflows/benchmark_tests.yml"
+      - "pyproject.toml"
+
+  pull_request:
+    branches:
+      - main
+    paths:
+      - "src/lerobot/envs/**"
+      - "src/lerobot/scripts/lerobot_eval.py"
+      - "docker/Dockerfile.benchmark.*"
+      - ".github/workflows/benchmark_tests.yml"
+      - "pyproject.toml"
+
+permissions:
+  contents: read
+
+env:
+  UV_VERSION: "0.8.0"
+  PYTHON_VERSION: "3.12"
+
+# Cancel in-flight runs for the same branch/PR.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  # ── LIBERO ────────────────────────────────────────────────────────────────
+  # Isolated image: lerobot[libero] only (hf-libero, dm-control, mujoco chain)
+  libero-integration-test:
+    name: Libero — build image + 1-episode eval
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+          lfs: true
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          cache-binary: false
+
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+
+      # Build the benchmark-specific image. The Dockerfile separates dep-install
+      # from source-copy, so code-only changes skip the slow uv-sync layer
+      # when the runner has a warm Docker daemon cache.
+      - name: Build Libero benchmark image
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          context: .
+          file: docker/Dockerfile.benchmark.libero
+          push: false
+          load: true
+          tags: lerobot-benchmark-libero:ci
+
+      - name: Run Libero smoke eval (1 episode)
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          # Named container (no --rm) so we can docker cp artifacts out.
+          # Output to /tmp inside the container — /artifacts doesn't exist
+          # and user_lerobot cannot create root-level dirs.
+          docker run --name libero-eval --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            lerobot-benchmark-libero:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              lerobot-eval \
+                --policy.path=pepijn223/smolvla_libero \
+                --env.type=libero \
+                --env.task=libero_spatial \
+                --eval.batch_size=1 \
+                --eval.n_episodes=1 \
+                --eval.use_async_envs=false \
+                --policy.device=cuda \
+                '--env.camera_name_mapping={\"agentview_image\": \"camera1\", \"robot0_eye_in_hand_image\": \"camera2\"}' \
+                --policy.empty_cameras=1 \
+                --output_dir=/tmp/eval-artifacts
+              python scripts/ci/extract_task_descriptions.py \
+                --env libero --task libero_spatial \
+                --output /tmp/eval-artifacts/task_descriptions.json
+            "
+
+      - name: Copy Libero artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/libero-artifacts
+          docker cp libero-eval:/tmp/eval-artifacts/. /tmp/libero-artifacts/ 2>/dev/null || true
+          docker rm -f libero-eval || true
+
+      - name: Parse Libero eval metrics
+        if: always()
+        run: |
+          python3 scripts/ci/parse_eval_metrics.py \
+            --artifacts-dir /tmp/libero-artifacts \
+            --env libero \
+            --task libero_spatial \
+            --policy pepijn223/smolvla_libero
+
+      - name: Upload Libero rollout video
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: libero-rollout-video
+          path: /tmp/libero-artifacts/videos/
+          if-no-files-found: warn
+
+      - name: Upload Libero eval metrics
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: libero-metrics
+          path: /tmp/libero-artifacts/metrics.json
+          if-no-files-found: warn
+
+      # ── LIBERO TRAIN+EVAL SMOKE ──────────────────────────────────────────────
+      # Train SmolVLA for 1 step (batch_size=1, dataset episode 0 only) then
+      # immediately runs eval inside the training loop (eval_freq=1, 1 episode).
+      # Tests the full train→eval-within-training pipeline end-to-end.
+      - name: Run Libero train+eval smoke (1 step, eval_freq=1)
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          docker run --name libero-train-smoke --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            lerobot-benchmark-libero:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              accelerate launch --num_processes=1 \$(which lerobot-train) \
+                --policy.path=lerobot/smolvla_base \
+                --policy.load_vlm_weights=true \
+                --policy.scheduler_decay_steps=25000 \
+                --policy.freeze_vision_encoder=false \
+                --policy.train_expert_only=false \
+                --dataset.repo_id=lerobot/libero \
+                --dataset.episodes=[0] \
+                --dataset.use_imagenet_stats=false \
+                --env.type=libero \
+                --env.task=libero_spatial \
+                '--env.camera_name_mapping={\"agentview_image\": \"camera1\", \"robot0_eye_in_hand_image\": \"camera2\"}' \
+                --policy.empty_cameras=1 \
+                --output_dir=/tmp/train-smoke \
+                --steps=1 \
+                --batch_size=1 \
+                --eval_freq=1 \
+                --eval.n_episodes=1 \
+                --eval.batch_size=1 \
+                --eval.use_async_envs=false \
+                --save_freq=1 \
+                --policy.push_to_hub=false \
+                '--rename_map={\"observation.images.image\": \"observation.images.camera1\", \"observation.images.image2\": \"observation.images.camera2\"}'
+            "
+
+      - name: Copy Libero train-smoke artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/libero-train-smoke-artifacts
+          docker cp libero-train-smoke:/tmp/train-smoke/. /tmp/libero-train-smoke-artifacts/ 2>/dev/null || true
+          docker rm -f libero-train-smoke || true
+
+      - name: Upload Libero train-smoke eval video
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: libero-train-smoke-video
+          path: /tmp/libero-train-smoke-artifacts/eval/
+          if-no-files-found: warn
+
+  # ── METAWORLD ─────────────────────────────────────────────────────────────
+  # Isolated image: lerobot[metaworld] only (metaworld==3.0.0, mujoco>=3 chain)
+  metaworld-integration-test:
+    name: MetaWorld — build image + 1-episode eval
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+          lfs: true
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          cache-binary: false
+
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+
+      - name: Build MetaWorld benchmark image
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          context: .
+          file: docker/Dockerfile.benchmark.metaworld
+          push: false
+          load: true
+          tags: lerobot-benchmark-metaworld:ci
+
+      - name: Run MetaWorld smoke eval (1 episode)
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          docker run --name metaworld-eval --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            lerobot-benchmark-metaworld:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              lerobot-eval \
+                --policy.path=pepijn223/smolvla_metaworld \
+                --env.type=metaworld \
+                --env.task=metaworld-push-v3 \
+                --eval.batch_size=1 \
+                --eval.n_episodes=1 \
+                --eval.use_async_envs=false \
+                --policy.device=cuda \
+                '--rename_map={\"observation.image\": \"observation.images.camera1\"}' \
+                --policy.empty_cameras=2 \
+                --output_dir=/tmp/eval-artifacts
+              python scripts/ci/extract_task_descriptions.py \
+                --env metaworld --task metaworld-push-v3 \
+                --output /tmp/eval-artifacts/task_descriptions.json
+            "
+
+      - name: Copy MetaWorld artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/metaworld-artifacts
+          docker cp metaworld-eval:/tmp/eval-artifacts/. /tmp/metaworld-artifacts/ 2>/dev/null || true
+          docker rm -f metaworld-eval || true
+
+      - name: Parse MetaWorld eval metrics
+        if: always()
+        run: |
+          python3 scripts/ci/parse_eval_metrics.py \
+            --artifacts-dir /tmp/metaworld-artifacts \
+            --env metaworld \
+            --task metaworld-push-v3 \
+            --policy pepijn223/smolvla_metaworld
+
+      - name: Upload MetaWorld rollout video
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: metaworld-rollout-video
+          path: /tmp/metaworld-artifacts/videos/
+          if-no-files-found: warn
+
+      - name: Upload MetaWorld eval metrics
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: metaworld-metrics
+          path: /tmp/metaworld-artifacts/metrics.json
+          if-no-files-found: warn
diff --git a/docker/Dockerfile.benchmark.libero b/docker/Dockerfile.benchmark.libero
new file mode 100644
index 000000000..620088b8b
--- /dev/null
+++ b/docker/Dockerfile.benchmark.libero
@@ -0,0 +1,42 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Benchmark image for LIBERO integration tests.
+# Extends the nightly GPU image (which already has all extras installed)
+# with the PR's source code and LIBERO-specific asset setup.
+#
+# Build:  docker build -f docker/Dockerfile.benchmark.libero -t lerobot-benchmark-libero .
+# Run:    docker run --gpus all --rm lerobot-benchmark-libero lerobot-eval ...
+
+FROM huggingface/lerobot-gpu:latest
+
+# Pre-download lerobot/libero-assets from HF Hub so nothing is fetched at
+# runtime (which times out on CI). Point the libero config at the cached path.
+# libero/libero/__init__.py calls input() when ~/.libero/config.yaml is missing,
+# so we write the config before any libero import can happen.
+RUN LIBERO_DIR=$(python -c \
+      "import importlib.util, os; s=importlib.util.find_spec('libero'); \
+       print(os.path.join(os.path.dirname(s.origin), 'libero'))") && \
+    mkdir -p /home/user_lerobot/.libero && \
+    python -c "\
+from huggingface_hub import snapshot_download; \
+snapshot_download(repo_id='lerobot/libero-assets', repo_type='dataset', \
+                  local_dir='/home/user_lerobot/.libero/assets')" && \
+    printf "assets: /home/user_lerobot/.libero/assets\nbddl_files: ${LIBERO_DIR}/bddl_files\ndatasets: ${LIBERO_DIR}/../datasets\ninit_states: ${LIBERO_DIR}/init_files\n" \
+    > /home/user_lerobot/.libero/config.yaml
+
+# Overlay the PR's source code on top of the nightly image.
+COPY --chown=user_lerobot:user_lerobot . .
+
+CMD ["/bin/bash"]
diff --git a/docker/Dockerfile.benchmark.metaworld b/docker/Dockerfile.benchmark.metaworld
new file mode 100644
index 000000000..96d9e89f9
--- /dev/null
+++ b/docker/Dockerfile.benchmark.metaworld
@@ -0,0 +1,27 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Benchmark image for MetaWorld integration tests.
+# Extends the nightly GPU image (which already has all extras installed)
+# with the PR's source code.
+#
+# Build:  docker build -f docker/Dockerfile.benchmark.metaworld -t lerobot-benchmark-metaworld .
+# Run:    docker run --gpus all --rm lerobot-benchmark-metaworld lerobot-eval ...
+
+FROM huggingface/lerobot-gpu:latest
+
+# Overlay the PR's source code on top of the nightly image.
+COPY --chown=user_lerobot:user_lerobot . .
+
+CMD ["/bin/bash"]
diff --git a/scripts/ci/extract_task_descriptions.py b/scripts/ci/extract_task_descriptions.py
new file mode 100644
index 000000000..5fbc1c35a
--- /dev/null
+++ b/scripts/ci/extract_task_descriptions.py
@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Extract natural-language task descriptions for a benchmark suite.
+
+Runs inside the benchmark Docker container (where the env library is installed)
+immediately after lerobot-eval, writing a JSON file that parse_eval_metrics.py
+picks up and embeds in metrics.json.
+
+Output format: {"<suite>_<task_idx>": "<nl instruction>", ...}
+
+Usage:
+    python scripts/ci/extract_task_descriptions.py \\
+        --env libero --task libero_spatial \\
+        --output /tmp/eval-artifacts/task_descriptions.json
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+
+def _libero_descriptions(task_suite: str) -> dict[str, str]:
+    from libero.libero import benchmark  # type: ignore[import-untyped]
+
+    suite_dict = benchmark.get_benchmark_dict()
+    if task_suite not in suite_dict:
+        print(
+            f"[extract_task_descriptions] Unknown LIBERO suite '{task_suite}'. "
+            f"Available: {list(suite_dict.keys())}",
+            file=sys.stderr,
+        )
+        return {}
+    suite = suite_dict[task_suite]()
+    return {f"{task_suite}_{i}": suite.get_task(i).language for i in range(suite.n_tasks)}
+
+
+def _metaworld_descriptions(task_name: str) -> dict[str, str]:
+    # MetaWorld tasks don't expose a separate NL description attribute;
+    # use a cleaned version of the task name as the description.
+    label = task_name.removeprefix("metaworld-").replace("-", " ").strip()
+    return {f"{task_name}_0": label}
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--env", required=True, help="Environment family (libero, metaworld, ...)")
+    parser.add_argument("--task", required=True, help="Task/suite name (e.g. libero_spatial)")
+    parser.add_argument("--output", required=True, help="Path to write task_descriptions.json")
+    args = parser.parse_args()
+
+    descriptions: dict[str, str] = {}
+    try:
+        if args.env == "libero":
+            descriptions = _libero_descriptions(args.task)
+        elif args.env == "metaworld":
+            descriptions = _metaworld_descriptions(args.task)
+        else:
+            print(
+                f"[extract_task_descriptions] No description extractor for env '{args.env}'.",
+                file=sys.stderr,
+            )
+    except Exception as exc:
+        print(f"[extract_task_descriptions] Warning: {exc}", file=sys.stderr)
+
+    out_path = Path(args.output)
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    out_path.write_text(json.dumps(descriptions, indent=2))
+    print(f"[extract_task_descriptions] {len(descriptions)} descriptions → {out_path}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/scripts/ci/parse_eval_metrics.py b/scripts/ci/parse_eval_metrics.py
new file mode 100644
index 000000000..897d9e81b
--- /dev/null
+++ b/scripts/ci/parse_eval_metrics.py
@@ -0,0 +1,147 @@
+#!/usr/bin/env python3
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Parse lerobot-eval output into a small metrics.json artifact.
+
+Reads eval_info.json written by lerobot-eval --output_dir and extracts the
+key metrics needed by the health dashboard. Handles both single-task and
+multi-task eval output formats.
+
+NOTE: This script runs on the bare CI runner (not inside Docker), so it
+must use only Python stdlib modules. Do not add third-party imports.
+
+Usage:
+    python scripts/ci/parse_eval_metrics.py \\
+        --artifacts-dir /tmp/libero-artifacts \\
+        --env libero \\
+        --task libero_spatial \\
+        --policy pepijn223/smolvla_libero
+
+Writes <artifacts-dir>/metrics.json. The CI workflow then uploads this file
+as a GitHub Actions artifact named "<env>-metrics".
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import sys
+from pathlib import Path
+
+
+def _safe_float(v: float | int | None) -> float | None:
+    if v is None:
+        return None
+    f = float(v)
+    return None if math.isnan(f) else f
+
+
+def _safe_int(v: float | int | None) -> int | None:
+    if v is None:
+        return None
+    f = float(v)
+    return None if math.isnan(f) else int(f)
+
+
+def _extract_metrics(info: dict) -> tuple[float | None, int | None, float | None, float | None]:
+    """Extract (pc_success, n_episodes, avg_sum_reward, eval_s) from eval_info.json.
+
+    Handles two output shapes:
+      - Single-task: {"aggregated": {"pc_success": 80.0, ...}}
+      - Multi-task:  {"overall": {"pc_success": 80.0, "n_episodes": 5, ...}}
+    """
+    for key in ("aggregated", "overall"):
+        if key not in info:
+            continue
+        agg = info[key]
+        pc = agg.get("pc_success")
+        n = agg.get("n_episodes")
+        reward = agg.get("avg_sum_reward")
+        eval_s = agg.get("eval_s")
+
+        if pc is not None and not math.isnan(pc):
+            return (
+                float(pc),
+                _safe_int(n),
+                _safe_float(reward),
+                _safe_float(eval_s),
+            )
+
+    return None, None, None, None
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    parser.add_argument("--artifacts-dir", required=True, help="Path to the mounted artifacts volume")
+    parser.add_argument("--env", required=True, help="Environment name (e.g. libero)")
+    parser.add_argument("--task", required=True, help="Task name (e.g. libero_spatial)")
+    parser.add_argument("--policy", required=True, help="Policy hub path (e.g. pepijn223/smolvla_libero)")
+    args = parser.parse_args()
+
+    artifacts_dir = Path(args.artifacts_dir)
+    eval_info_path = artifacts_dir / "eval_info.json"
+
+    pc_success: float | None = None
+    n_episodes: int | None = None
+    avg_sum_reward: float | None = None
+    eval_s: float | None = None
+
+    if eval_info_path.exists():
+        try:
+            info = json.loads(eval_info_path.read_text())
+            pc_success, n_episodes, avg_sum_reward, eval_s = _extract_metrics(info)
+        except (json.JSONDecodeError, KeyError, TypeError) as exc:
+            print(f"[parse_eval_metrics] Warning: could not parse eval_info.json: {exc}", file=sys.stderr)
+    else:
+        print(
+            f"[parse_eval_metrics] Warning: {eval_info_path} not found — eval may have failed.",
+            file=sys.stderr,
+        )
+
+    task_descriptions: dict[str, str] = {}
+    task_desc_path = artifacts_dir / "task_descriptions.json"
+    if task_desc_path.exists():
+        try:
+            task_descriptions = json.loads(task_desc_path.read_text())
+        except json.JSONDecodeError as exc:
+            print(
+                f"[parse_eval_metrics] Warning: could not parse task_descriptions.json: {exc}",
+                file=sys.stderr,
+            )
+
+    metrics = {
+        "env": args.env,
+        "task": args.task,
+        "policy": args.policy,
+        "pc_success": pc_success,
+        "n_episodes": n_episodes,
+        "avg_sum_reward": avg_sum_reward,
+        "eval_s": eval_s,
+        "task_descriptions": task_descriptions,
+    }
+
+    out_path = artifacts_dir / "metrics.json"
+    out_path.write_text(json.dumps(metrics, indent=2))
+    print(f"[parse_eval_metrics] Written: {out_path}")
+    print(json.dumps(metrics, indent=2))
+
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())