evaluation
Guide VLA-Adapter LIBERO and CALVIN evaluation, command generation, outputs, prerequisites, metrics, and benchmark cautions.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- evaluation
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Guide VLA-Adapter LIBERO and CALVIN evaluation, command generation, outputs, prerequisites, metrics, and benchmark cautions.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/benchmarks.md5.3 KB
- 参考文档references/evaluation-workflows.md11.0 KB
- 参考文档references/troubleshooting.md6.2 KB
- 脚本scripts/build_eval_command.py8.9 KB
03 SKILL.md 正文
Evaluation (external-checkout adapter)
This sub-skill renders plans for native LIBERO/CALVIN evaluators in a separate
checkout. It contains only documentation and a command renderer: it does not
contain experiments/robot/ or vla-scripts/, does not run an evaluation, and
is not a self-contained benchmark runtime.
Set the absolute source root and enter it before any native workflow:
export VLA_ADAPTER_REPO_ROOT=/abs/path/to/VLA-Adapter
cd "$VLA_ADAPTER_REPO_ROOT"
python -m pip install -e "$VLA_ADAPTER_REPO_ROOT"
Native entrypoints are experiments/robot/libero/run_libero_eval.py and
vla-scripts/evaluate_calvin.py. The bundled builder only prints commands
beginning with cd <absolute-repo-root> &&; it never imports the model or
launches a rollout. External prerequisites are CUDA-capable PyTorch, the base
VLA-Adapter package, a complete compatible checkpoint, the selected LIBERO or
CALVIN checkout and assets/data/config layout, and a working EGL/OpenGL/MuJoCo
or pybullet renderer. Video runs may additionally require imageio/ffmpeg.
Read first
- Read references/evaluation-workflows.md for checkpoint-to-command workflows, expected outputs, and benchmark-specific runtime behavior.
- Read references/benchmarks.md for suite names, checkpoint names, default evaluation volumes, and published LIBERO/CALVIN metrics.
- Read references/troubleshooting.md for dependency, renderer, checkpoint, unnormalization-key, log, and video failures.
- Run
python "$VLA_ADAPTER_SKILL_ROOT/sub-skills/evaluation/scripts/build_eval_command.py" --repo-root "$VLA_ADAPTER_REPO_ROOT" --helpto inspect the non-executing command renderer.
Operating route
- Identify the benchmark: use LIBERO for
libero_spatial,libero_object,libero_goal, orlibero_10; use CALVIN forcalvin_abc. - Confirm the checkpoint family and suite match. Prefer the Pro checkpoints for published reproduction unless the user explicitly asks for the original release or a custom checkpoint.
- Confirm downstream prerequisites before launching: CUDA-capable PyTorch, the base VLA-Adapter package dependencies, the external benchmark stack, benchmark assets, and the checkpoint files.
- Generate the command with the bundled builder, then review the resulting
CUDA_VISIBLE_DEVICES=... python ...invocation before execution. - Capture stdout/stderr into an
eval_logs/file for long runs. LIBERO also writes internal text logs and rollout videos; CALVIN writes evaluation result files and per-subtask videos. - Interpret only full-volume runs against published metrics: LIBERO defaults to 500 episodes per suite, while CALVIN reports average successful sequence length across 1,000 five-instruction sequences.
Critical cautions
- LIBERO evaluation depends on the repository's image preprocessing: both third-person and wrist camera images are rotated 180 degrees before policy input.
- LIBERO/OpenVLA action execution normalizes the gripper action to
[-1, +1], binarizes it, then inverts the OpenVLA gripper sign for environment execution. - CALVIN reports average sequence length, not a single binary episode success rate; also inspect chain success rates for 1 through 5 consecutive instructions.
- External LIBERO and CALVIN stacks were not part of the base package import surface. Treat them as downstream prerequisites, not guaranteed installed dependencies.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).