Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/harbor/harbor.py", line 171, in _split_generators
                  raise DataFilesNotFoundError("No task.toml or instruction.md files found")
              datasets.exceptions.DataFilesNotFoundError: No task.toml or instruction.md files found
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

MobileManipBench (MagicSim large-object tasks): results and progress

Updated 2026-11-05 01:15 UTC (generated, do not edit). 20 tasks × 3 robots (go2plus, stretch3, sawyer_fin) = 60 task–robot pairs; each pair has Easy / Medium / Hard + DR = 180 level cells.

怎么重建这个页面、监控进度的脚本(给别的 agent 看):整个文件夹 rebuild_kit/,从 00_START_HERE.md 开始。页面的版式照 yyoraa/mobilemanipbench-results。

Progress

finished tasks (all 3 robots: Easy/Medium/Hard success, DR ≥ 5 successful draws, owner validated) 0 / 20
task–robot pairs finished (simulation stages done + owner validated) 0 / 60
task–robot pairs with all stages done, waiting for the owner 17 / 60
task–robot pairs with Easy, Medium and Hard all successful 18 / 60
task–robot pairs with DR ≥ 5 successful draws 30 / 60
level cells (Easy / Medium / Hard) with a real success 85 / 180
complete records (success + sim_info, replayable without physics) / all successful records 372 / 2340
task–robot pairs with ≥ 5 complete records 24 / 60

A success is the environment's own full-success termination (no teleport, no weld, no scripted motion; NavMesh + DWB base, MagicSim command chain). A pair is finished only after the owner validated it. There are no videos in this project: stage 3 is the domain-randomisation draws (object instance swap + object pose DR, the base must really move). (The raw numbers are in INDEX.md and status_table/.)

需要你看一下(卡住的格子)

规则:一个格子在当前阶段累计活跃满 2 小时还没过,agent 就不再自己反复调,把失败原因和证据交给你看,你看了告诉我们问题出在哪里,agent 同时去做下一个格子。下表列出已满 2 小时、仍有 kit 在跑、当前阶段还没过的格子和它最近一次失败的原因:

Agent Task Robot 当前阶段 累计活跃 这格跑过 最近一次失败的原因
a03 F012 go2plus 阶段一 Easy 约 9.9 小时 17 次 episode ended (env truncated by a skill) during phase nav_to_B
a03 F012 sawyer_fin 阶段一 Easy 约 9.3 小时 52 次 skill place_approach ended with truncated=3 (failed: global planner failed to plan)

现在谁在做什么

只列有 kit 正在运行的格子(一个 kit 就是一次仿真任务,由显卡上的进程判断)。"当前阶段"是表里第一个还没通过的阶段;"正在跑"是这个 kit 跑的难度(DR = 域随机化的一次抽样)。"这格已待"是从这个任务、这个机器人的第一份运行记录算起的时间,"累计活跃"是把相邻运行记录的间隔(单段最多算 1 小时)加起来的估计,不是 GPU 时间。 一个 agent 同时可以在几个格子上有 kit;没有 kit 的 agent 写"暂无 kit 在跑"(它可能在改代码或等待显存)。

Agent Task Robot 当前阶段 正在跑 这格的首次运行(美国中部) 这格已待 累计活跃 最近一次运行记录
a03 F012 送来接收盒,再拔挡片让球落入 go2plus 阶段一 Easy DR 11-03 20:44 22.5 小时 约 7.3 小时 2.2 小时前
a03 F012 送来接收盒,再拔挡片让球落入 sawyer_fin 阶段一 Easy DR 11-04 13:32 5.7 小时 约 5.7 小时 19 分钟前
a01 — 暂无 kit 在跑
a02 — 暂无 kit 在跑
a04 — 暂无 kit 在跑

每个格子用了多久

上图画的就是上面"现在谁在做什么"那张表里的数字:条长 = 这个格子的第一份运行记录到现在("这格已待"),括号里是"累计活跃"。柱子的颜色是它现在所处的阶段。 下图和下面的表是一个阶段从开始到完成用了多久,从 11-02 00:37(美国中部)开始记录,这之前完成的阶段没有记录;第一次被看到时还在阶段一的格子,从它的第一份运行记录算起,其余的从记录开始算起。 时间分辨率约 10 分钟,约每 10 分钟更新一次。"活跃"只算有运行记录的时间,agent 想问题、改代码的时间不在里面。

每个格子每个阶段用了多久

完成时间(美国中部) Agent Task Robot 完成的阶段 用了 其中活跃
11-03 02:34 a02 F006 sawyer_fin 阶段一 Easy 10.8 小时 53 分钟
11-02 19:02 a04 F016 stretch3 阶段一 Easy 10 分钟 0 分钟
11-02 18:30 a04 F016 go2plus 阶段一 Easy 43.8 小时 7.5 小时
11-02 17:37 a01 F003 go2plus 阶段二 Medium+Hard 10 分钟 10 分钟
11-02 17:05 a02 F006 stretch3 阶段一 Easy 10 分钟 10 分钟
11-02 17:05 a04 F016 stretch3 阶段一 Easy 15.4 小时 53 分钟
11-02 16:33 a01 F005 go2plus 阶段一 Easy 10 分钟 0 分钟
11-02 16:33 a01 F003 stretch3 阶段二 Medium+Hard 10 分钟 10 分钟
11-02 16:22 a01 F004 go2plus 阶段一 Easy 10.3 天 3.1 天
11-02 16:11 a01 F003 go2plus 阶段一 Easy 14 分钟 3 分钟
11-02 16:11 a01 F001 stretch3 阶段二 Medium+Hard 14 分钟 25 分钟
11-02 15:57 a01 F002 go2plus 阶段一 Easy 31 分钟 20 分钟
11-02 15:46 F003 stretch3 阶段二 Medium+Hard 10 分钟 0 分钟
11-02 15:36 a01 F002 stretch3 阶段二 Medium+Hard 10 分钟 10 分钟
11-02 15:26 F001 stretch3 阶段二 Medium+Hard 10 分钟 0 分钟
11-02 15:26 a01 F001 go2plus 阶段一 Easy 10.3 天 3.0 天
11-02 15:15 a01 F004 stretch3 阶段一 Easy 10 分钟 0 分钟
11-02 15:04 F002 stretch3 阶段一 Easy 10 分钟 0 分钟
11-02 13:08 a04 F019 stretch3 阶段三 DR 12.5 小时 7 分钟
11-02 11:04 a04 F016 sawyer_fin 阶段三 DR 8.2 小时 31 分钟

总表

每个机器人下面有五列:**Easy、Medium、Hard、DR(阶段三,蓝底)、人工验证(紫底)**。阶段一 = Easy 成功;阶段二 = Medium 和 Hard 都成功;阶段三 = DR 至少 5 次成功;人工验证只由你来标。

符号 出现在哪 意思
✅ Easy、Medium、Hard 这个难度最近一次标准运行(不带 DR)环境判定完全成功
✅ s/N DR N 次 DR 抽样里 s 次成功,s ≥ 5,阶段三完成
✅ 人工 你已经验证通过,这一对(任务 × 机器人)才算完成
❌ Easy、Medium、Hard 跑过,但最近一次没成功
❌ s/N DR 跑过 N 次抽样只成功 s 次(还不到 5 次)
· 任意列 还没有
绿色加粗 ▶ a02 运行 · 已3.1小时 正在跑的那一列 这个 agent 的 kit 正在这个难度(或 DR)上跑;后面是这个格子已经待了多久(和"现在谁在做什么"表里一致)

格子下面的小字是该结果出现的时间(美国中部时间,月-日 时:分)。

TaskAgentgo2plusstretch3sawyer_fin
EasyMediumHardDR
阶段三
人工
验证
EasyMediumHardDR
阶段三
人工
验证
EasyMediumHardDR
阶段三
人工
验证
F001
把移动白板推到展示位并转向观众
a01✅
11-02 15:24
✅
11-02 17:09
❌
11-02 19:12
✅ 30/79
10-26 22:17
·✅
11-02 13:14
✅
11-02 15:24
✅
11-02 16:06
✅ 136/212
10-23 07:35
·✅
10-22 15:42
✅
10-22 17:34
✅
10-22 19:21
✅ 10/10
10-23 02:29
·
F002
沿屏风外缘移动,把两折屏风展开成L形
a01✅
11-02 15:50
❌
11-02 17:24
❌
11-02 19:26
✅ 32/80
10-26 20:18
·✅
11-02 14:54
✅
11-02 15:31
✅
11-02 16:18
✅ 160/198
10-23 07:42
·✅
10-22 15:06
✅
10-22 17:59
✅
10-22 19:29
✅ 10/10
10-23 02:37
·
F003
拉出长桌延伸板并停在终点
a01✅
11-02 16:04
✅
11-02 17:33
✅
11-02 19:33
✅ 44/75
10-23 15:00
·✅
11-02 13:29
✅
11-02 15:36
✅
11-02 16:22
✅ 184/270
10-24 07:46
·✅
10-22 15:18
✅
10-22 18:08
✅
10-22 19:39
✅ 10/10
10-23 02:47
·
F004
把轻型拼桌推到另一张桌边
a01✅
11-02 16:14
❌
11-02 18:24
❌
11-02 20:24
❌ 0/81·✅
11-02 15:05
✅
11-02 15:43
❌
11-02 16:48
✅ 51/280
10-24 00:27
·✅
10-22 15:28
✅
10-22 18:32
✅
10-22 16:57
✅ 10/10
10-23 03:22
·
F005
把宽箱盖滑到旁侧支架,露出箱口
a01✅
11-02 16:27
✅
11-02 18:38
❌
11-02 20:39
✅ 11/79
10-28 18:58
·❌
11-02 15:09
✅
11-02 15:50
✅
11-02 16:56
✅ 239/268
10-23 07:58
·✅
10-22 15:48
✅
10-22 18:42
✅
10-23 03:36
✅ 10/10
10-23 03:27
·
F006
取一块长板,在两个支座之间搭桥
a02❌
10-26 17:58
❌
10-26 07:41
❌
10-26 18:59
❌ 0/127·✅
10-23 04:28
✅
10-26 03:25
❌
11-03 13:13
✅ 121/278
10-26 13:14
·✅
11-03 02:28
✅
10-22 15:28
✅
10-22 15:45
✅ 206/221
10-26 05:16
·
F007
送大漏斗到接料位,再倒入三颗球
a02❌
11-03 12:56
❌
10-26 09:17
❌
10-25 04:57
❌ 0/5·❌
11-03 12:21
❌
11-01 01:53
❌
11-01 02:02
❌ 0/26·❌
11-03 12:51
❌
11-01 01:57
❌
11-01 02:09
❌ 1/52·
F008
找到指定大画框并斜靠到展示架
a02❌
11-03 13:07
❌
10-26 09:26
❌
10-25 05:37
❌ 0/20·❌
11-03 12:46
❌
11-01 02:55
❌
11-01 03:23
❌ 0/26·❌
11-03 12:56
❌
11-01 02:19
❌
11-01 02:26
❌ 0/53·
F009
把空料箱送到传送带末端接料
a02❌
11-03 14:29
❌
11-03 13:38
❌
11-03 14:04
❌ 2/151·❌
10-26 11:14
❌
10-26 04:30
❌
10-26 04:34
❌ 0/206·✅
10-22 18:39
✅
10-23 01:56
✅
10-26 05:42
✅ 215/228
10-26 04:12
·
F010
把展台上的大防尘布揭到后侧
a02❌
11-03 12:16
❌
10-25 21:55
❌
10-25 22:42
❌ 0/5·❌
11-03 12:12
❌
11-01 03:37
❌
11-01 04:07
❌ 0/26·❌
11-03 11:50
❌
11-01 02:54
❌
11-01 02:58
❌ 0/61·
F011
补上多米诺缺口,再推倒短链
a03✅
11-03 20:49
❌
11-04 14:28
❌
11-04 16:20
❌ 0/47·✅
11-03 22:30
❌
11-04 10:07
❌
11-04 12:55
✅ 58/197
11-04 00:09
·✅
11-04 13:07
❌
11-04 15:58
❌
11-04 15:43
✅ 49/96
11-04 14:16
·
F012
送来接收盒,再拔挡片让球落入
a03❌
11-03 22:52
··❌ 0/11
▶ a03 运行 · 已22.5 小时
·✅
11-03 09:39
❌
11-03 13:31
❌
11-03 16:38
❌ 2/38·❌
11-04 15:30
··❌ 1/27
▶ a03 运行 · 已5.7 小时
·
F013
舀起几颗玩具豆,跨台倒进碗里
a03❌
11-04 07:55
··❌ 0/17·❌
11-03 12:17
··❌ 0/25·❌
11-04 00:19
··❌ 0/22·
F014
搬球压下跷跷板,让另一端抬起
a03✅
11-03 17:45
·❌
10-30 21:58
✅ 29/46
11-03 21:36
·✅
11-04 01:08
✅
11-04 01:23
✅
11-04 04:26
✅ 250/335
11-04 02:00
·✅
11-04 17:16
✅
11-04 16:26
✅
11-04 16:13
✅ 95/95
11-04 16:28
·
F015
取大圈,圈住两颗球并推到收集区
a03❌
11-03 21:40
··❌ 0/22·❌
11-03 19:47
··✅ 14/137
11-03 23:38
·✅
11-04 12:51
✅
11-04 13:35
✅
11-04 15:23
✅ 67/106
11-04 13:37
·
F016
把落地风扇移到指定位置并朝向人
a04✅
11-02 18:28
❌
11-03 02:05
❌
11-03 02:38
❌ 1/7·✅
11-02 18:55
❌
11-02 20:28
❌
11-02 20:33
❌ 4/7·❌
11-02 16:28
✅
11-02 18:48
✅
11-02 19:06
❌ 4/7·
F017
拉空行李箱,按人的停/继续手势送达
a04❌
11-02 19:02
✅
11-02 22:05
✅
11-02 22:32
❌ 2/5·✅
11-02 17:52
✅
11-02 17:59
✅
11-02 18:05
✅ 7/7
11-02 21:19
·✅
11-02 20:03
✅
11-02 17:36
✅
11-02 17:46
✅ 7/7
11-02 21:03
·
F018
携宽空托盘让人先通过,再送到收纳架
a04❌
11-02 19:31
❌
11-03 02:57
❌
11-03 03:17
❌ 0/7·✅
11-02 17:31
✅
11-02 18:14
❌
11-02 18:27
✅ 7/7
11-02 22:42
·✅
11-02 20:22
✅
11-02 19:58
❌
11-02 20:17
❌ 2/7·
F019
记住人选的图案,再把双面大展牌转到正确面
a04✅
11-02 20:31
❌
11-02 23:23
❌
11-02 23:36
❌ 0/7·✅
11-02 17:12
✅
11-02 19:16
✅
11-02 19:41
❌ 3/7·✅
11-02 17:17
✅
11-02 17:55
❌
11-02 17:59
✅ 7/7
11-02 23:40
·
F020
绕到大转盘另一侧,把指定展品转到人面前
a04❌
11-02 21:25
❌
11-03 00:41
❌
11-03 01:22
❌ 3/7·✅
11-02 17:27
✅
11-02 19:51
✅
11-02 20:09
✅ 6/6
11-03 02:38
·✅
11-02 17:25
❌
11-02 18:23
❌
11-02 18:36
✅ 7/7
11-03 00:29
·

Where things are

  • INDEX.md / index.csv: one row per task × robot with the result, the agent, Easy/Medium/Hard, DR, human validation, run count and the last failure reason.
  • by_agent/aNN.md: the agent's five tasks on the three robots. Agent a01 has F001–F005, a02 F006–F010, a03 F011–F015, a04 F016–F020.
  • cells/<task>_<title>_<agent>/<robot>/STATUS.md: that pair's results in words (all 60 folders always exist).
  • rebuild_kit/: everything needed to rebuild this page: 00_START_HERE.md, FRONT_PAGE_SPEC.md, HOW_TO_REBUILD.md, scripts/ (the table builder, the page generator, the chart drawer).
  • status_table/: the live progress table (csv, json), the input of the page generator.

About the records

The simulation records themselves (per-step json, camera frames) are not in this repo; they are on the owner's private Hugging Face backup. This page only reports the results of the runs. A path never changes when a pair is validated: the validation state lives in INDEX.md.

Downloads last month
6,322