[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M78SMXC8DWSVYCKF12SEJYAB":3,"skill-same:01M78SMXC8DWSVYCKF12SEJYAB":60},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":26,"commitSha":27,"skillDirectory":28,"forksCount":29,"repositoryHtmlUrl":30,"repositoryHomepageUrl":6,"ownerHtmlUrl":31,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":32,"install":33,"resources":45},"01M78SMXC8DWSVYCKF12SEJYAB","agibot-conversion",null,"Routes AgiBotWorld raw-tree conversion requests to LeRobot, covering end-effector selection, task filtering, episode alignment, depth and video handling, execution controls, and recovery.","VectorSpaceLab\u002FAREX-Skill","VectorSpaceLab","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F180637102?v=4","01M7538C5WPSJGWF7KHR4EFZ9F","01M7538C6R2GFP53NPXTCZNBRT",351,"MIT","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Fany4lerobot\u002Fsub-skills\u002Fagibot-conversion\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: \"agibot-conversion\"\ndescription: \"Routes AgiBotWorld raw-tree conversion requests to LeRobot,\n  covering end-effector selection, task filtering, episode alignment, depth and\n  video handling, execution controls, and recovery.\"\ndisable-model-invocation: true\nmetadata:\n  disco-role: operating\nlicense: MIT\n---\n\n# AgiBotWorld conversion\n\nUse this route when a user needs to convert an AgiBotWorld\u002FAgiBot-World raw\nrelease into a LeRobot dataset, select gripper\u002Fdexhand\u002Ftactile tasks, preserve\nAgiBot state and action names, or diagnose an AgiBot conversion failure. This\nroute is dataset-specific. Load [`generic-conversion`](..\u002Fgeneric-conversion\u002FSKILL.md)\nfor shared adapter lifecycle, temporary aggregation, local\u002FRay executor\nsemantics, resume logs, cleanup, and Hub publication; do not duplicate that\npipeline here.\n\n## Safety and compatibility boundary\n\n- Treat the raw tree and the output root as separate resources. Inventory both\n  before scheduling work, and never use the source tree as an output root.\n- Prefer a read-only structural preflight, `--executor local`, and `--debug`\n  for the first run. The default executor is Ray, but Ray is optional for a\n  single-node conversion.\n- Do not enable `--push-to-hub` until the local dataset opens and its task,\n  episode, feature, and modality counts have been reviewed. Require an\n  explicit `--repo-id` and confirm the destination is not an unintended\n  existing dataset.\n- The AgiBot implementation uses a custom metadata flush and dataset writer.\n  It is version-sensitive and must be checked against the installed LeRobot\n  API before conversion. In the verified environment, current LeRobot exposes\n  `LeRobotDataset` and `LeRobotDatasetMetadata` under\n  `lerobot.datasets.lerobot_dataset`, while the source import style also\n  expects exports from `lerobot.datasets` and `lerobot.datasets.dataset_writer`.\n  Treat an import-layout mismatch as a blocking compatibility issue, not as a\n  reason to edit metadata by guesswork.\n- Do not run a conversion, Ray cluster, video decode, download, or Hub push\n  while only planning or validating this route.\n\n## Route checklist\n\n1. Confirm the release root contains `task_info\u002F`, `observations\u002F`, and\n   `proprio_stats\u002F`; confirm that the requested task IDs and episode IDs are\n   present in the release rather than inferred from a filename alone.\n2. Choose exactly one `--eef-type`: `gripper`, `dexhand`, or `tactile`. Keep\n   task families separate. The gripper route means the tasks not reserved for\n   dexhand or tactile in the built-in mapping; it is not a universal “all\n   tasks” switch.\n3. Check the selected task JSON records, HDF5 state\u002Faction paths, every required\n   RGB\u002Ftactile video, and depth frame count when `--save-depth` is requested.\n   Use the schemas in [data formats](references\u002Fdata-formats.md).\n4. Decide task granularity with `--episodes-per-task`. Use `1` for load\n   balancing and isolation, or a larger value when task-start overhead and\n   memory are acceptable. Chunk names are deterministic but are not semantic\n   task IDs.\n5. Select execution and resource flags from [workflows](references\u002Fworkflows.md),\n   including `--executor`, `--cpus-per-task`, `--tasks-per-job`, `--workers`,\n   `--resume-dir`, and `--debug`. Start locally; only use Ray after cluster\n   ownership, shared paths, optional dependencies, and worker memory have been\n   confirmed. Budget roughly 20 GiB per active task and about 3 CPU cores per\n   task as a starting point, then reduce concurrency if memory is constrained.\n6. Preserve the frame task label (`task_name | init_scene_text`) and the\n   per-episode `action_config`. Do not flatten or rename nested state\u002Faction\n   keys without recording an explicit downstream schema change.\n7. Interpret failures conservatively: missing video is skipped before loading,\n   corrupt video is skipped when episode saving fails, action longer than state\n   skips the episode, and depth count mismatch must be investigated rather than\n   padded silently. See [troubleshooting](references\u002Ftroubleshooting.md).\n8. Validate the local result before resume or publication: feature names and\n   shapes, task labels, action metadata, episode counts, depth presence, video\n   readability, and absence of unexpected temporary output. Then hand shared\n   cleanup, aggregation, resume, and Hub behavior to `generic-conversion`.\n\n## AgiBot-specific transformation\n\nThe converter reads state and action arrays from `proprio_stats.h5`, emits one\nframe per state sample, and attaches the natural-language task label to every\nframe. Empty action arrays become zero-valued frames of the configured shape.\nShort action arrays are reconstructed from the corresponding state stream and\nplaced at the recorded action indices when those indices exist; an empty\n`action\u002F...\u002Findex` may fall back to the joint index. An action stream longer\nthan the state stream is treated as corrupt and the episode is skipped. This is\nalignment policy, not generic padding and must be recorded in provenance.\n\nRGB and tactile streams remain video features; optional head depth is loaded as\nfloat32 image data from depth frames and scaled from millimetres to metres. The\ncustom writer stores task indices, action configuration, episode statistics,\nand video references while writing episode metadata as Parquet. These details\nare summarized, with no source-code dependency, in the bundled references.\n\n## Difficult synthetic case\n\nFor a mixed raw tree containing one gripper task, one dexhand task, and one\nshort-action episode, first filter by `--eef-type` and explicit `--task-ids`,\nthen verify that the short action is reconstructed only when the index\u002Fstate\ncontract is valid. Add a task with a missing depth frame and another with a\nmissing or corrupt video: the video case should be skipped with a reason; the\ndepth case should stop or be excluded by preflight rather than silently\ninventing frames. This case is suitable for an integration usability test and\nrequires no real AgiBot data.\n\n## Handoff and evidence\n\nReport the source release shape, selected end-effector and tasks, skipped\nepisodes with reasons, alignment and depth decisions, execution\u002Fresource\nflags, compatibility checks, output validation, and any unresolved optional\nbackend gap. Evidence for this route was distilled from the repository README,\nthe AgiBot conversion README, `agibot_h5.py`, `agibot_utils\u002Fagibot_utils.py`,\n`agibot_utils\u002Fconfig.py`, and `agibot_utils\u002Flerobot_utils.py`; these artifact\nnames are provenance only and are not runtime dependencies.\n","# AgiBotWorld conversion\n\nUse this route when a user needs to convert an AgiBotWorld\u002FAgiBot-World raw\nrelease into a LeRobot dataset, select gripper\u002Fdexhand\u002Ftactile tasks, preserve\nAgiBot state and action names, or diagnose an AgiBot conversion failure. This\nroute is dataset-specific. Load [`generic-conversion`](..\u002Fgeneric-conversion\u002FSKILL.md)\nfor shared adapter lifecycle, temporary aggregation, local\u002FRay executor\nsemantics, resume logs, cleanup, and Hub publication; do not duplicate that\npipeline here.\n\n## Safety and compatibility boundary\n\n- Treat the raw tree and the output root as separate resources. Inventory both\n  before scheduling work, and never use the source tree as an output root.\n- Prefer a read-only structural preflight, `--executor local`, and `--debug`\n  for the first run. The default executor is Ray, but Ray is optional for a\n  single-node conversion.\n- Do not enable `--push-to-hub` until the local dataset opens and its task,\n  episode, feature, and modality counts have been reviewed. Require an\n  explicit `--repo-id` and confirm the destination is not an unintended\n  existing dataset.\n- The AgiBot implementation uses a custom metadata flush and dataset writer.\n  It is version-sensitive and must be checked against the installed LeRobot\n  API before conversion. In the verified environment, current LeRobot exposes\n  `LeRobotDataset` and `LeRobotDatasetMetadata` under\n  `lerobot.datasets.lerobot_dataset`, while the source import style also\n  expects exports from `lerobot.datasets` and `lerobot.datasets.dataset_writer`.\n  Treat an import-layout mismatch as a blocking compatibility issue, not as a\n  reason to edit metadata by guesswork.\n- Do not run a conversion, Ray cluster, video decode, download, or Hub push\n  while only planning or validating this route.\n\n## Route checklist\n\n1. Confirm the release root contains `task_info\u002F`, `observations\u002F`, and\n   `proprio_stats\u002F`; confirm that the requested task IDs and episode IDs are\n   present in the release rather than inferred from a filename alone.\n2. Choose exactly one `--eef-type`: `gripper`, `dexhand`, or `tactile`. Keep\n   task families separate. The gripper route means the tasks not reserved for\n   dexhand or tactile in the built-in mapping; it is not a universal “all\n   tasks” switch.\n3. Check the selected task JSON records, HDF5 state\u002Faction paths, every required\n   RGB\u002Ftactile video, and depth frame count when `--save-depth` is requested.\n   Use the schemas in [data formats](references\u002Fdata-formats.md).\n4. Decide task granularity with `--episodes-per-task`. Use `1` for load\n   balancing and isolation, or a larger value when task-start overhead and\n   memory are acceptable. Chunk names are deterministic but are not semantic\n   task IDs.\n5. Select execution and resource flags from [workflows](references\u002Fworkflows.md),\n   including `--executor`, `--cpus-per-task`, `--tasks-per-job`, `--workers`,\n   `--resume-dir`, and `--debug`. Start locally; only use Ray after cluster\n   ownership, shared paths, optional dependencies, and worker memory have been\n   confirmed. Budget roughly 20 GiB per active task and about 3 CPU cores per\n   task as a starting point, then reduce concurrency if memory is constrained.\n6. Preserve the frame task label (`task_name | init_scene_text`) and the\n   per-episode `action_config`. Do not flatten or rename nested state\u002Faction\n   keys without recording an explicit downstream schema change.\n7. Interpret failures conservatively: missing video is skipped before loading,\n   corrupt video is skipped when episode saving fails, action longer than state\n   skips the episode, and depth count mismatch must be investigated rather than\n   padded silently. See [troubleshooting](references\u002Ftroubleshooting.md).\n8. Validate the local result before resume or publication: feature names and\n   shapes, task labels, action metadata, episode counts, depth presence, video\n   readability, and absence of unexpected temporary output. Then hand shared\n   cleanup, aggregation, resume, and Hub behavior to `generic-conversion`.\n\n## AgiBot-specific transformation\n\nThe converter reads state and action arrays from `proprio_stats.h5`, emits one\nframe per state sample, and attaches the natural-language task label to every\nframe. Empty action arrays become zero-valued frames of the configured shape.\nShort action arrays are reconstructed from the corresponding state stream and\nplaced at the recorded action indices when those indices exist; an empty\n`action\u002F...\u002Findex` may fall back to the joint index. An action stream longer\nthan the state stream is treated as corrupt and the episode is skipped. This is\nalignment policy, not generic padding and must be recorded in provenance.\n\nRGB and tactile streams remain video features; optional head depth is loaded as\nfloat32 image data from depth frames and scaled from millimetres to metres. The\ncustom writer stores task indices, action configuration, episode statistics,\nand video references while writing episode metadata as Parquet. These details\nare summarized, with no source-code dependency, in the bundled references.\n\n## Difficult synthetic case\n\nFor a mixed raw tree containing one gripper task, one dexhand task, and one\nshort-action episode, first filter by `--eef-type` and explicit `--task-ids`,\nthen verify that the short action is reconstructed only when the index\u002Fstate\ncontract is valid. Add a task with a missing depth frame and another with a\nmissing or corrupt video: the video case should be skipped with a reason; the\ndepth case should stop or be excluded by preflight rather than silently\ninventing frames. This case is suitable for an integration usability test and\nrequires no real AgiBot data.\n\n## Handoff and evidence\n\nReport the source release shape, selected end-effector and tasks, skipped\nepisodes with reasons, alignment and depth decisions, execution\u002Fresource\nflags, compatibility checks, output validation, and any unresolved optional\nbackend gap. Evidence for this route was distilled from the repository README,\nthe AgiBot conversion README, `agibot_h5.py`, `agibot_utils\u002Fagibot_utils.py`,\n`agibot_utils\u002Fconfig.py`, and `agibot_utils\u002Flerobot_utils.py`; these artifact\nnames are provenance only and are not runtime dependencies.\n",{"name":5,"license":14,"metadata":23,"description":7,"disable-model-invocation":25},{"disco-role":24},"operating",true,"main","ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6","skills\u002Frepositories\u002Frepo-skills\u002Fany4lerobot\u002Fsub-skills\u002Fagibot-conversion",27,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab","2026-09-03 10:25:21",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":25},[35,38,41],{"tool":36,"command":37},"npx","npx skills add VectorSpaceLab\u002FAREX-Skill --skill agibot-conversion",{"tool":39,"command":40},"bunx","bunx skills add VectorSpaceLab\u002FAREX-Skill --skill agibot-conversion",{"tool":42,"command":43},"pnpm","pnpm dlx skills add VectorSpaceLab\u002FAREX-Skill --skill agibot-conversion","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Fany4lerobot\u002Fsub-skills\u002Fagibot-conversion\u002FSKILL.md",[46,52,56],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"references\u002Fdata-formats.md","REFERENCE","参考文档",8214,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Fany4lerobot\u002Fsub-skills\u002Fagibot-conversion\u002Freferences\u002Fdata-formats.md",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"references\u002Ftroubleshooting.md",10210,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Fany4lerobot\u002Fsub-skills\u002Fagibot-conversion\u002Freferences\u002Ftroubleshooting.md",{"path":57,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":58,"sourceHtmlUrl":59},"references\u002Fworkflows.md",7745,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Fany4lerobot\u002Fsub-skills\u002Fagibot-conversion\u002Freferences\u002Fworkflows.md",[61,69,76,83,90,97],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":65,"licenseFileUrl":6,"sourceHtmlUrl":66,"categories":67,"keywords":68,"publishedAt":18},"01M799NYSH1R3RVAHSM2D88Y84","passnet-triton-opt","Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block\u002Fgrid\u002Fwarp 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.","Apache-2.0","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-triton-opt\u002FSKILL.md",[],[],{"skillNo":70,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":65,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"publishedAt":18},"01M799NS5APEC0N1FZ7XZVW6BN","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.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-solve\u002FSKILL.md",[],[],{"skillNo":77,"name":78,"nameZh":6,"description":79,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":65,"licenseFileUrl":6,"sourceHtmlUrl":80,"categories":81,"keywords":82,"publishedAt":18},"01M799NNY9BKSM21V5R7QVTBJX","passnet-skill","PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under .\u002Fpass_dir\u002F, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-skill\u002FSKILL.md",[],[],{"skillNo":84,"name":85,"nameZh":6,"description":86,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":65,"licenseFileUrl":6,"sourceHtmlUrl":87,"categories":88,"keywords":89,"publishedAt":18},"01M799NHH6NP7Z1RK77KGXAN6S","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()\u002Freplacement_args()\u002Freplacement_func() files, when a pass \"failed to match\", or when deciding how to fuse multiple kernels\u002Fops.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-pattern-fusion\u002FSKILL.md",[],[],{"skillNo":91,"name":92,"nameZh":6,"description":93,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":65,"licenseFileUrl":6,"sourceHtmlUrl":94,"categories":95,"keywords":96,"publishedAt":18},"01M799ND2J4WNC2PGAZ0A06NDF","passnet-orchestrate","Round\u002Fbatch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep\u002Frevert 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.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-orchestrate\u002FSKILL.md",[],[],{"skillNo":98,"name":99,"nameZh":6,"description":100,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":65,"licenseFileUrl":6,"sourceHtmlUrl":101,"categories":102,"keywords":103,"publishedAt":18},"01M799N8SBTMY02ZDG2AZF5ZWE","passnet-feedback","Fast, deterministic feedback for PassNet work: pre-flight pattern\u002Fmatch 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).","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-feedback\u002FSKILL.md",[],[]]