| # 可复现评测全集 — 使用说明 (REPRODUCE) |
|
|
| 本仓库是 GR00T N1.7 / pi0 / Genie-Envisioner 在 ManiSkill 上的 pairwise-OOD、 |
| all-factor、conflict、VLM 评测的**可复现全集**(除 checkpoint 权重外的全部: |
| 代码、仿真、评测结果、复现步骤)。 |
|
|
| ## 目录结构 |
| ``` |
| . |
| ├── README.md # 全部实验结果总表(7 节)+ 附录:干扰项/口径明细 |
| ├── result.md # GR00T pair-grid + all_factor 结果(精简版) |
| ├── REPRODUCE.md # 本文件:怎么跑 |
| ├── code/ # 全部评测代码(client + runner + dispatcher) |
| ├── eval_results/ # 每个 ckpt 的逐 seed 明细 txt/jsonl |
| │ ├── results_gr00t_pair_*/ # GR00T pair-grid 硬口径(+color_size_easy) |
| │ ├── results_af/ # GR00T all_factor full-factor |
| │ ├── results_pi0_pairwise/ # pi0 pair-grid |
| │ ├── conflict_env_{gr00t,genie}/ # conflict 双轴 env success |
| │ └── vlm_eval/ # Gemini-2.5-flash VLM 判定 + FDR_summary |
| └── simulation/ |
| ├── eval_simulation_src.tar.gz # ManiSkill 仿真(VerbObjectColor-v1 环境) |
| ├── Maniskill_gen_new.tar.gz # 数据采集定义(get_env_id_and_color 等) |
| └── VERSION.txt # 仿真 git commit |
| ``` |
|
|
| ## 0. 准备 |
|
|
| ### 0.1 仿真就位(代码已是文件夹,可在线浏览;仅二进制 3D 资源是 tar) |
| ```bash |
| # 代码已在 simulation/eval_simulation/ 与 simulation/Maniskill_gen_new/(文件夹,可直接看) |
| # 把二进制资源(网格/贴图/hdr,382 个)解开,覆盖回 eval_simulation/ 即补全可跑: |
| tar xzf simulation/eval_simulation_ASSETS.tar.gz -C simulation/ # 解出 eval_simulation/.../*.stl,*.glb,*.hdr ... |
| # 然后把整个 simulation/eval_simulation 与 simulation/Maniskill_gen_new 放到 /workspace/ 下 |
| cp -r simulation/eval_simulation /workspace/ |
| cp -r simulation/Maniskill_gen_new /workspace/ |
| ``` |
| > 说明:代码(.py/.urdf/.xml/.mtl/.json/.sh,864 文件 7.9M)放成文件夹便于浏览/复现; |
| > 二进制 3D 资源(.stl/.glb/.obj/.png/.hdr 等 382 文件 110M)单独 tar,因为它们本就不可读、 |
| > 且文件数多会撞 HF 提交限流。解压后 eval_simulation 即与原始一致(1246 文件)。 |
| |
| ### 0.2 venv(三套,互相隔离) |
| - `/workspace/groot_eval/.venv_ms` ManiSkill 物理/渲染 + 评测 client(numpy/torch/mani_skill/zmq) |
| - `/workspace/groot_eval/.venv_groot` GR00T N1.7 推理服务器(gr00t 包 + transformers) |
| - `/venv/pi0_eval` pi0 openpi 推理服务器(仅 pi0 评测需要) |
|
|
| ### 0.3 checkpoint(本仓库**不含**权重,从 HF 拉) |
| - pair-grid(f6/f12/f18):`yqi19/gr00t_public_pair_ckpt` → 放到 `/workspace/gr00t_pair_ckpt/<ckpt>/` |
| - all_factor(f50/n*):`yqi19/gr00t_all_factor_evaluation` → `/workspace/groot_eval/gr00t_af_ckpts/<ckpt>/checkpoint-10000/` |
| - pi0:`/workspace/pi0_ckpt/<ckpt>/17999/` |
|
|
| ## 1. GR00T pair-grid(硬口径,3-seed) |
| 单卡单 ckpt(内部跑 seed 42/40/41,起一次 GR00T zmq server): |
| ```bash |
| cd /workspace/groot_eval |
| SEEDS_OVERRIDE="42 40 41" bash code/run_groot_one_ckpt.sh <ckpt_name> <gpu> <port> |
| # 例: ... color_spatial_stair_f18 0 5700 |
| # color_size EASY 口径: 额外 COLOR_SIZE_MODE=easy |
| ``` |
| - 实现:`code/groot_grid_eval.py`(env 构建 + HARD/EASY 干扰项 + 评分) |
| - 输出:`results_gr00t_pair_<experiment>/<ckpt>/gr00t_<exp>_<ckpt>_seed<sd>.txt` |
| - 实验由 ckpt 名前缀解析(color_size/color_spatial/verb_spatial/spatial_size/spatial_object/verb_size) |
|
|
| ### 多 ckpt 并行(8 卡常驻调度器) |
| ```bash |
| tmux new-session -d -s gpair_dispatcher "bash code/gpair_dispatcher.sh" |
| # 把 "<ckpt>:<seed>[:easy]" 逐行写进 logs/gr00t_pair/vs_queue.txt 即可; |
| # all_factor 用 "af:<ckpt>" 写进 af_queue.txt |
| ``` |
|
|
| ## 2. GR00T all_factor(full-factor,对齐 pi0.5) |
| ```bash |
| bash code/run_af_one_ckpt_fast.sh <af_ckpt> <gpu> <port> |
| # 内部:run_full_factor_groot.sh + groot_full_factor_main.py,seeds 40/41/42, |
| # sample_n=200, 200ep, max_steps=500, no_distractor=0.70 |
| # 输出: results_af/<ckpt>/full_factor_<ckpt>_seed<sd>.txt |
| ``` |
| |
| ## 3. pi0 pair-grid |
| ```bash |
| SEEDS_OVERRIDE="42 40 41" bash code/run_pi0_one_ckpt.sh <ckpt> <gpu> <port> |
| # sim_backend=cpu(pi0 server 与 GPU-sim 冲突);实现 code/pi0_grid_eval.py |
| # 输出: results_pi0_pairwise/<ckpt>/pi0_<exp>_<ckpt>_seed<sd>.txt |
| ``` |
| |
| ## 4. conflict-experiment(双轴 env success) |
| ```bash |
| # GR00T: groot_main.py + run_ood_groot_inference.sh(起 GR00T server) |
| # 输出 overall_<factorA>_success / overall_<factorB>_success |
| ``` |
| |
| ## 5. VLM 判定(Gemini-2.5-flash)+ FDR |
| - 用 Gemini-2.5-flash 看 conflict 视频判 factor_followed,产出 `vlm_eval/<model>/vlm_eval_<exp>.jsonl` |
| - FDR=(S_f1−S_f2)/(S_f1+S_f2),三方对照见 `eval_results/vlm_eval/FDR_summary.txt` |
| |
| ## 6. 干扰项 / 口径明细(可复现关键) |
| **见 `README.md` 末尾「附:可复现配置」** —— 每个实验的 target/distractor 尺寸、 |
| spatial 锚点坐标、num_distractors、指令模板、task_difficulty / max_steps / replan |
| 全部列清。照着即可逐 cell 复现。 |
| |
| ## 7. 结果速览 |
| 最强:color_spatial_stair_f18(GR00T 51.1% / pi0 54.9%);含 size 的任务(color_size/verb_size) |
| 对 VLA 最难(~8–35%)。完整 73 ckpt × 3-seed 表见 `README.md`。 |
| |
