evaluation_all / result.md
yqi19's picture
Upload folder using huggingface_hub
8aa2acf verified
|
Raw
History Blame Contribute Delete
11.7 kB

仿真评测结果总汇 (GR00T N1.7 / Genie-Envisioner)

生成时间:2026-05-25 19:20:58 UTC

本文件汇总全部实验:① GR00T pair-grid 硬/软口径 ② GR00T all_factor full-factor ③ conflict 双轴 env success(GR00T vs Genie)④ VLM(Gemini-2.5-flash)FDR 对照。 数值为成功率 %;pair-grid/all_factor 均为 3-seed 平均。


一、GR00T pair-grid(硬口径,3-seed)

color_size (HARD)

ckpt s42 s40 s41 3-seed 均
color_size_Lrandom_f12 24.1 27.3 24.5 25.3%
color_size_Lrandom_f18 35.6 34.7 34.3 34.8%
color_size_Lrandom_f6 7.9 10.2 6.9 8.3%
color_size_random_f12 13.4 15.3 11.6 13.4%
color_size_random_f18 27.3 32.9 28.2 29.4%
color_size_random_f6 9.7 9.3 10.6 9.8%
color_size_stair_f12 28.2 28.2 28.2 28.2%
color_size_stair_f18 26.9 27.8 25.9 26.8%
color_size_stair_f6 7.4 8.3 6.0 7.2%

color_size (EASY)

ckpt s42 s40 s41 3-seed 均
color_size_Lrandom_f12 23.6 24.5 22.7 23.6%
color_size_Lrandom_f18 36.6 38.0 37.0 37.2%
color_size_Lrandom_f6 5.1 10.6 9.3 8.3%
color_size_random_f12 14.4 14.8 16.2 15.1%
color_size_random_f18 28.7 33.3 31.0 31.0%
color_size_random_f6 6.9 11.1 6.9 8.3%
color_size_stair_f12 29.2 27.8 29.6 28.8%
color_size_stair_f18 26.4 25.5 27.8 26.5%
color_size_stair_f6 9.3 8.3 6.5 8.0%

color_spatial

ckpt s42 s40 s41 3-seed 均
color_spatial_Lrandom_f18 52.4 42.9 46.7 47.3%
color_spatial_random_f12 33.3 27.6 33.8 31.5%
color_spatial_random_f18 42.4 36.7 41.0 40.0%
color_spatial_random_f6 23.3 26.2 24.8 24.7%
color_spatial_stair_f18 54.8 45.7 52.9 51.1%
color_spatial_stair_f6 24.3 19.5 21.4 21.7%

verb_spatial

ckpt s42 s40 s41 3-seed 均
verb_spatial_Lrandom_f12 31.9 29.5 32.4 31.2%
verb_spatial_Lrandom_f18 32.9 30.0 34.8 32.5%
verb_spatial_Lrandom_f6 22.4 31.0 27.1 26.8%
verb_spatial_random_f12 31.4 31.9 37.6 33.6%
verb_spatial_random_f18 26.2 24.3 23.3 24.6%
verb_spatial_random_f6 16.2 16.2 14.3 15.5%
verb_spatial_stair_f12 26.7 28.6 25.2 26.8%
verb_spatial_stair_f18 34.3 37.1 35.2 35.5%
verb_spatial_stair_f6 23.3 27.1 22.9 24.4%

spatial_size

ckpt s42 s40 s41 3-seed 均
spatial_size_Lrandom_f12 38.1 30.0 33.3 33.8%
spatial_size_Lrandom_f18 32.4 35.7 33.8 33.9%
spatial_size_Lrandom_f6 24.3 21.0 24.8 23.3%
spatial_size_random_f12 37.6 33.3 34.3 35.0%
spatial_size_random_f18 27.6 30.5 28.1 28.7%
spatial_size_random_f6 21.4 17.6 24.3 21.1%
spatial_size_stair_f12 36.2 37.6 40.5 38.1%
spatial_size_stair_f18 46.7 46.7 51.0 48.1%

spatial_object

ckpt s42 s40 s41 3-seed 均
spatial_object_Lrandom_f12 40.0 40.0 38.6 39.5%
spatial_object_Lrandom_f18 34.8 31.4 32.9 33.0%
spatial_object_random_f12 27.1 32.4 29.5 29.6%
spatial_object_random_f18 34.3 36.7 32.4 34.4%
spatial_object_random_f6 17.6 17.6 15.7 16.9%
spatial_object_stair_f12 30.5 39.0 34.8 34.7%
spatial_object_stair_f18 32.9 34.8 28.6 32.1%
spatial_object_stair_f6 28.6 32.9 26.2 29.2%

verb_size

ckpt s42 s40 s41 3-seed 均
verb_size_Lrandom_f12 13.4 15.3 16.7 15.1%
verb_size_Lrandom_f18 22.2 23.6 23.6 23.1%
verb_size_Lrandom_f6 8.3 9.3 7.4 8.3%
verb_size_random_f12 23.1 24.5 25.0 24.2%
verb_size_random_f18 14.4 13.9 11.6 13.3%
verb_size_random_f6 10.2 8.8 7.4 8.8%
verb_size_stair1_f12 20.4 17.6 17.1 18.3%
verb_size_stair1_f18 17.6 16.2 18.5 17.4%

二、GR00T all_factor (full-factor, pi0.5 对齐, 3-seed)

ckpt s42 s40 s41 3-seed 均
all_factor_Lrandom_f50_n100 46.5 43.5 49.5 46.5%
all_factor_Lrandom_f50_n200 44.0 43.0 36.0 41.0%
all_factor_Lrandom_f50_n400 45.5 48.0 47.5 47.0%
all_factor_Lrandom_n200 34.0 26.0 35.0 31.6%
all_factor_Lrandom_n400 31.0 33.0 37.0 33.6%
all_factor_Lrandom_n800 32.0 31.5 38.0 33.8%
all_factor_stairoriginal2_n200 38.5 35.0 38.5 37.3%
all_factor_stairverbsize_n200 36.5 32.5 38.0 35.6%
all_factor_stairverbsize_n400 44.5 41.5 38.5 41.5%
all_factor_stairverbsize_n800 29.0 31.5 32.0 30.8%
all_factor_stairvss_f50_n200 40.0 40.5 44.5 41.6%
all_factor_stairvss_f50_n400 42.0 43.0 43.5 42.8%
all_factor_stairvss_f50_n800 42.5 45.5 51.5 46.5%
all_factor_true_random_n200 36.0 31.0 34.5 33.8%
all_factor_true_random_n400 40.5 38.0 36.0 38.1%
all_factor_true_random_n800 41.5 38.5 44.0 41.3%

三、Conflict-experiment 双轴 env success (seed42, 200 ep)

每个实验两个因子的 success(模型把任务做对、且作用在正确因子目标上)。

experiment GR00T 因子A GR00T 因子B Genie 因子A Genie 因子B
color_object 15.0% 14.0% 11.7% 6.7%
color_size 47.0% 29.5% 1.5% 0.0%
color_spatial 5.5% 39.5% 10.5% 38.5%
size_object 18.5% 29.5% 0.0% 0.0%
spatial_object 9.0% 22.0% 1.5% 16.2%
spatial_size 1.0% 25.0% 30.5% 18.5%
verb_color 24.5% 24.0% 13.5% 11.5%
verb_object 8.0% 31.5% 4.9% 3.2%
verb_size 17.0% 9.0% 19.5% 22.5%
verb_spatial 38.5% 4.0% 7.0% 14.5%

四、VLM (Gemini-2.5-flash) FDR 对照

============================================================================================
Factor Dominance Rate (FDR) 三方对照 — FDR=(S_f1-S_f2)/(S_f1+S_f2), >0偏因子A <0偏因子B
env  = 基于 ManiSkill 环境 success(每模型 200 episodes)
vlm  = 基于 Gemini-2.5-flash 看视频判 factor_followed(每模型 400 视频/实验)
============================================================================================
experiment     A/B           env_GR00T env_genie  vlm_GR00T  vlm_genie
--------------------------------------------------------------------------------------------
color_object   color/shape      +0.034    +0.034     +0.228     +0.222
color_size     color/size       +0.229    +1.000     +0.524     +0.309
color_spatial  color/spati      -0.756    -0.571     +0.084     +0.365
size_object    size/shape       -0.229    +0.602     -0.241     -0.114
spatial_object spati/shape      -0.419    -0.713     -0.095     -0.209
spatial_size   spati/size       -0.923    +0.245     +0.110     +0.237
verb_color     verb/color       +0.010    +0.080     -0.068     -0.220
verb_object    verb/shape       -0.595    -0.062     -0.253     -0.250
verb_size      verb/size        +0.308    -0.071     +0.251     +0.326
verb_spatial   verb/spati       +0.812    -0.349     -0.122     -0.142
--------------------------------------------------------------------------------------------

VLM 判定明细(N_f1 / N_f2 / neither|error,共400):

五、pi0 (openpi pi-zero) pair-grid (硬口径, 3-seed)

与 GR00T pair-grid 同口径(同色同形干扰项),但 sim_backend=cpu(pi0 server+GPU-sim 冲突)。注意 pi0 与 GR00T 推理栈不同,数值横向比 GR00T 仅供参考。

pi0 — color_spatial

ckpt s42 s40 s41 3-seed 均
color_spatial_Lrandom_f12 52.4 41.4 49.5 47.7%
color_spatial_Lrandom_f18 53.8 41.0 48.6 47.8%
color_spatial_Lrandom_f6 28.6 23.3 26.7 26.2%
color_spatial_random_f12 38.6 32.4 37.6 36.2%
color_spatial_random_f18 51.0 41.9 47.6 46.8%
color_spatial_random_f6 27.6 21.4 23.3 24.1%
color_spatial_stair_f12 49.0 38.1 42.4 43.1%
color_spatial_stair_f18 60.0 49.0 55.7 54.9%
color_spatial_stair_f6 31.0 24.8 28.6 28.1%

pi0 — color_size

ckpt s42 s40 s41 3-seed 均
color_size_Lrandom_f12 27.3 - - 1/3(部分)
color_size_Lrandom_f18 30.6 29.2 27.3 29.0%
color_size_Lrandom_f6 26.4 - - 1/3(部分)
color_size_random_f12 - - - 0/3(部分)
color_size_random_f18 28.7 25.9 23.1 25.9%
color_size_random_f6 - - - 0/3(部分)
color_size_stair_f18 26.4 28.2 24.1 26.2%

附:可复现配置(干扰项 / 口径明细)

公共设置(pair-grid 全实验)

  • 环境:VerbObjectColor-v1(ManiSkill),obs=rgb,control=pd_joint_pos
  • task_difficulty=1.5,sim/render_backend=gpu,max_episode_steps=150,replan_steps=10
  • 每 seed:cells × reps,target≈210 episodes;seeds={42,40,41}(3-seed 平均)
  • 干扰项核心:同色(且多数同形 cube)干扰项,强制模型靠目标因子区分,而非靠颜色/形状走捷径
  • spatial 锚点(xy,米)+ 抖动 ±0.012:left(-0.10,0) right(0.10,0) middle(0,0) front(0,0.10) behind(0,-0.10)

color_size — HARD 口径

  • verb=lift, shape=cube;指令 "Lift the {size} {color} cube."
  • 干扰项 = 同色 cube,仅尺寸不同。(target_scale, [干扰scale...], 干扰数):
    • small (0.72,[1.20],1) · large (1.34,[0.80],1) · smaller (0.82,[1.08],1) · larger (1.18,[0.92],1) · smallest (0.78,[1.00,1.24],2) · largest (1.26,[1.00,0.80],2)

color_size — EASY 口径

  • 同 HARD 但干扰项恒为 1 个、对比拉大 ~2×:
    • small (0.65,[1.35],1) · large (1.40,[0.65],1) · smaller (0.70,[1.30],1) · larger (1.30,[0.70],1) · smallest (0.65,[1.40],1) · largest (1.40,[0.65],1)

color_spatial — HARD 口径

  • shape=cube,verb~{lift,grasp,push} 随机;指令 "{Verb} the {color} cube {spatial_phrase}."
  • 干扰项 = 1 个同色 cube,放在不同 spatial 锚点;target_size_scale=1.0

verb_spatial — HARD 口径

  • shape=cube,color 每 cell 随机;指令 "{Verb} the cube {spatial_phrase}."
  • 干扰项 = 1 个同色 cube,放在不同 spatial 锚点

spatial_size — HARD 口径

  • shape=cube,color&verb 随机;指令 "{Verb} the {size} cube {spatial_phrase}."
  • 干扰项 = 1 个同色 cube,不同 spatial + 不同尺寸(目标偏小→干扰 scale=1.25;目标偏大→0.75)

spatial_object — HARD 口径

  • color&verb 随机;指令 "{Verb} the {shape} {spatial_phrase}."
  • 干扰项 = 1 个同色、不同形状物体,放在不同 spatial 锚点

verb_size — HARD 口径

  • shape=cube,color 随机;指令 "{Verb} the {size} cube."
  • 干扰项 = 同色 cube,仅尺寸不同(同 color_size HARD 的 HARD_SIZE 表)

all_factor — full-factor(对齐 pi0.5)

  • 单指令全因子采样;sample_n=200, sample_seed=42, 200 episodes/seed
  • max_episode_steps=500, no_distractor_prob=0.70, 默认 difficulty, sim=gpu;seeds={40,41,42}

pi0 pair-grid

  • 与 GR00T pair-grid 同口径,但 sim_backend=cpu(pi0 server 与 GPU-sim 冲突);其余同上