{"id":263,"date":"2026-10-29T18:30:00","date_gmt":"2026-10-29T13:00:00","guid":{"rendered":"https:\/\/www.graveiensai.com\/blog\/?p=263"},"modified":"2026-10-27T16:01:47","modified_gmt":"2026-10-27T10:31:47","slug":"collect-teleoperation-data","status":"publish","type":"post","link":"https:\/\/www.graveiensai.com\/blog\/collect-teleoperation-data\/","title":{"rendered":"How to Collect Teleoperation Data: A Pilot Scoping Playbook for Robot Demonstrations"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\"><i>Last updat\u00aded: Octo\u00adber 2026<\/i><\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Key takeaways<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Know\u00ading how to col\u00adlect tele\u00adop\u00ader\u00ada\u00adtion data starts before any ses\u00adsion: write the robot and task brief, then agree what a suc\u00adcess\u00adful episode looks like.<\/li>\n\n\n\n<li>A pilot should answer one ques\u00adtion: can this rig, these oper\u00ada\u00adtors and this QA process pro\u00adduce episodes your pol\u00adi\u00adcy can train on?<\/li>\n\n\n\n<li>Log more than video. Joint states, grip\u00adper com\u00admands, time\u00adstamps and episode out\u00adcomes decide whether demon\u00adstra\u00adtions are usable.<\/li>\n\n\n\n<li>Size the pilot by the deci\u00adsion it must sup\u00adport, not by the size of the even\u00adtu\u00adal dataset.<\/li>\n\n\n\n<li>Close every pilot with a writ\u00adten go\/no-go review against the accep\u00adtance cri\u00adte\u00adria you set at the start.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Robot time is expen\u00adsive, and a tele\u00adop\u00ader\u00ada\u00adtion cam\u00adpaign that starts with\u00adout a clear spec usu\u00adal\u00adly pro\u00adduces episodes nobody can train on. This play\u00adbook is for robot\u00adics, ML and pro\u00adgram leads who already know what tele\u00adop\u00ader\u00ada\u00adtion is and now need to run a first data pilot. If you need the basics first, our <a href=\"https:\/\/www.graveiensai.com\/blog\/teleoperation\/\">tele\u00adop\u00ader\u00ada\u00adtion explain\u00ader<\/a> cov\u00aders def\u00adi\u00adn\u00adi\u00adtions, con\u00adtrol modes and rig types. Here you get a five-phase plan for how to col\u00adlect tele\u00adop\u00ader\u00ada\u00adtion data in a pilot, a sam\u00adple brief, a log\u00adging spec, a ses\u00adsion-day run\u00adbook, a siz\u00ading method and a go\/no-go score\u00adcard.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><b>Ready to plan yours?<\/b> <a href=\"https:\/\/www.graveiensai.com\/contact-us\">Scope a pilot billed only on approved files.<\/a><\/p>\n\n\n\n<h2 class=\"wp-block-heading\">How do you collect teleoperation data for a robot learning pilot?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">To col\u00adlect tele\u00adop\u00ader\u00ada\u00adtion data, a trained oper\u00ada\u00adtor dri\u00adves the robot (or a matched rig) through a defined task while the sys\u00adtem records syn\u00adchro\u00adnized cam\u00adera streams, joint states and com\u00admands for every episode. Each episode is then checked against writ\u00adten accep\u00adtance cri\u00adte\u00adria. A pilot runs this loop at small scale to prove the rig, oper\u00ada\u00adtors and QA work before you com\u00admit to vol\u00adume.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In short: a pilot is a test of your data pipeline, not a small ver\u00adsion of your final dataset. Treat it as an exper\u00adi\u00adment with a hypoth\u00ade\u00adsis, a bud\u00adget and a deci\u00adsion date.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What should a teleoperation pilot prove?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A tele\u00adop\u00ader\u00ada\u00adtion pilot should prove three things: that oper\u00ada\u00adtors can com\u00adplete the task reli\u00adably on the cho\u00adsen rig, that every episode is logged in a for\u00admat your train\u00ading code can load, and that QA can sep\u00ada\u00adrate good episodes from bad ones using cri\u00adte\u00adria every\u00adone agreed in advance. If any of the three fails, scal\u00ading only mul\u00adti\u00adplies the prob\u00adlem.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Write the hypoth\u00ade\u00adsis down. For exam\u00adple: \u201cTwo trained oper\u00ada\u00adtors can pro\u00adduce accept\u00aded pick-and-place episodes on our arm at a steady rate, with every episode car\u00adry\u00ading syn\u00adchro\u00adnized wrist and over\u00adhead video, joint states and an out\u00adcome label.\u201d A pilot with a stat\u00aded hypoth\u00ade\u00adsis ends with a clear answer. A pilot with\u00adout one ends with a fold\u00ader of files.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">The five phases of a teleoperation data pilot<\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"720\" src=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__01-five-phases.webp\" alt=\"Five phases of a teleoperation data pilot: brief and criteria, match rig, train and run, log and QA, go\/no-go review.\" class=\"wp-image-283\" srcset=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__01-five-phases.webp 1600w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__01-five-phases-300x135.webp 300w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__01-five-phases-1024x461.webp 1024w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__01-five-phases-768x346.webp 768w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__01-five-phases-1536x691.webp 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\"><figcaption class=\"wp-element-caption\">The five phas\u00ades of a tele\u00adop\u00ader\u00ada\u00adtion data pilot.<\/figcaption><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">Phase 1: Write the brief and the acceptance criteria<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">The brief names the robot plat\u00adform (make, mod\u00adel, degrees of free\u00addom, end effec\u00adtor), the task list, the envi\u00adron\u00adment, safe\u00adty lim\u00adits and the tar\u00adget cycle time. Add the accep\u00adtance cri\u00adte\u00adria in the same doc\u00adu\u00adment: what counts as task suc\u00adcess, which fail\u00adures are allowed with\u00adin an episode (a regrasp, say), and which ones void it (a col\u00adli\u00adsion, a dropped object, a missed step).<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Agree the cri\u00adte\u00adria before the first ses\u00adsion. Cri\u00adte\u00adria writ\u00adten after footage exists invite dis\u00adputes, because every\u00adone is now judg\u00ading data they already paid for.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Phase 2: Match the rig to the robot and the task<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Most pilots choose between a leader-fol\u00adlow\u00ader arm pair and VR-based con\u00adtrol. Leader-fol\u00adlow\u00ader setups such as the open-source ALOHA sys\u00adtem suit fine biman\u00adu\u00adal manip\u00adu\u00adla\u00adtion, because the oper\u00ada\u00adtor feels the motion through a match\u00ading arm. VR-based sys\u00adtems such as Open-Tele\u00adVi\u00adsion stream the robot\u00ad\u2019s cam\u00adera view to a head\u00adset and suit humanoid plat\u00adforms and longer tasks. Our <a href=\"https:\/\/www.graveiensai.com\/blog\/teleoperation\/\">guide to tele\u00adop\u00ader\u00ada\u00adtion types<\/a> com\u00adpares the options in depth.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"800\" src=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__02-rig-families.webp\" alt=\"Two rig families compared: a leader-follower arm pair and a VR-headset-controlled robot arm.\" class=\"wp-image-284\" srcset=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__02-rig-families.webp 1600w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__02-rig-families-300x150.webp 300w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__02-rig-families-1024x512.webp 1024w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__02-rig-families-768x384.webp 768w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__02-rig-families-1536x768.webp 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\"><figcaption class=\"wp-element-caption\">Two com\u00admon rig fam\u00adi\u00adlies for demon\u00adstra\u00adtion cap\u00adture.<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">For the pilot, the ques\u00adtion is nar\u00adrow\u00ader: which rig lets an oper\u00ada\u00adtor fin\u00adish your task clean\u00adly at least most of the time? If you can\u00adnot decide, run a few ses\u00adsions on each and com\u00adpare accep\u00adtance rates.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Phase 3: Train operators and run supervised sessions<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Oper\u00ada\u00adtor skill dri\u00adves data qual\u00adi\u00adty. Plan prac\u00adtice ses\u00adsions that are record\u00aded but exclud\u00aded from the dataset, then super\u00advised ses\u00adsions where a lead flags bad episodes in real time. Rotate at least two oper\u00ada\u00adtors so you can see whether the data reflects the task or one per\u00adson\u00ad\u2019s habits.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Phase 4: Log the right signals and QA every episode<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Decide the log\u00adging spec before cap\u00adture. The table below is a prac\u00adti\u00adcal min\u00adi\u00admum for most manip\u00adu\u00adla\u00adtion pilots.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Sig\u00adnal<\/th><th>Why it mat\u00adters<\/th><th>What to spec\u00adi\u00adfy<\/th><\/tr><\/thead><tbody><tr><td>Cam\u00adera streams (wrist, over\u00adhead)<\/td><td>Poli\u00adcies learn from what the robot sees<\/td><td>Res\u00ado\u00adlu\u00adtion, frame rate, cam\u00adera IDs and mount\u00ading<\/td><\/tr><tr><td>Joint states and com\u00admands<\/td><td>The action labels your pol\u00adi\u00adcy imi\u00adtates<\/td><td>Rate, units and joint order<\/td><\/tr><tr><td>Grip\u00adper or hand state<\/td><td>Grasp tim\u00ading is often where poli\u00adcies fail<\/td><td>Open\/close val\u00adues or fin\u00adger posi\u00adtions<\/td><\/tr><tr><td>Time\u00adstamps on one clock<\/td><td>Mis\u00adaligned streams break action-obser\u00adva\u00adtion pairs<\/td><td>Clock source and allowed drift<\/td><\/tr><tr><td>Episode meta\u00adda\u00adta<\/td><td>Lets you fil\u00adter and rebal\u00adance lat\u00ader<\/td><td>Task ID, oper\u00ada\u00adtor ID, envi\u00adron\u00adment, out\u00adcome, fail\u00adure rea\u00adson<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">QA then checks each episode against the accep\u00adtance cri\u00adte\u00adria and records a rea\u00adson for every rejec\u00adtion. Rejec\u00adtion rea\u00adsons are the most use\u00adful out\u00adput of a pilot, because each one points to a spec, rig or train\u00ading change.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Phase 5: Deliver, review and decide<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Deliv\u00ader accept\u00aded episodes in the for\u00admat your train\u00ading stack already loads, togeth\u00ader with the rejec\u00adtion log. Then hold a go\/no-go review against the hypoth\u00ade\u00adsis from the start of the pilot.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">A sample teleoperation pilot brief<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A one-page brief is enough for most pilots. Fill in every field before the first ven\u00addor or inter\u00adnal plan\u00adning call; blanks are where scope creep starts.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Field<\/th><th>Exam\u00adple entry<\/th><\/tr><\/thead><tbody><tr><td>Hypoth\u00ade\u00adsis<\/td><td>Two oper\u00ada\u00adtors can pro\u00adduce accept\u00aded pick-and-place episodes at a steady rate on our arm<\/td><\/tr><tr><td>Plat\u00adform<\/td><td>Arm mod\u00adel, end effec\u00adtor, con\u00adtrol inter\u00adface and firmware ver\u00adsion, locked for the pilot<\/td><\/tr><tr><td>Tasks<\/td><td>Three tasks, each with num\u00adbered steps and allowed object posi\u00adtions<\/td><\/tr><tr><td>Accep\u00adtance cri\u00adte\u00adria<\/td><td>Task com\u00adplete, no col\u00adli\u00adsion, no dropped object, all streams present and in sync<\/td><\/tr><tr><td>Log\u00adging spec<\/td><td>Streams, rates, clock source, file for\u00admat and episode meta\u00adda\u00adta fields<\/td><\/tr><tr><td>Size and stop rule<\/td><td>Fixed batch per task; stop and review if accep\u00adtance stays low after ear\u00adly ses\u00adsions<\/td><\/tr><tr><td>Deci\u00adsion date<\/td><td>The day the go\/no-go review hap\u00adpens, with named atten\u00addees<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">Session-day runbook<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Most lost episodes trace back to some\u00adthing that could have been checked before the first record\u00ading of the day. A short run\u00adbook keeps ses\u00adsions con\u00adsis\u00adtent across oper\u00ada\u00adtors and days.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><b>Before the ses\u00adsion:<\/b> con\u00adfirm cal\u00adi\u00adbra\u00adtion files are cur\u00adrent, check clock sync across cam\u00aderas and robot, test the emer\u00adgency stop, and record a short test episode.<\/li>\n\n\n\n<li><b>Dur\u00ading the ses\u00adsion:<\/b> the lead logs task ID and oper\u00ada\u00adtor ID per episode, flags fail\u00adures with a rea\u00adson, and paus\u00ades for breaks before fatigue shows in the data.<\/li>\n\n\n\n<li><b>After the ses\u00adsion:<\/b> copy files to two loca\u00adtions, run an auto\u00admat\u00aded check for miss\u00ading streams and dropped frames, and file the day\u2019s rejec\u00adtion rea\u00adsons.<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">How many episodes should a teleoperation pilot include?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A tele\u00adop\u00ader\u00ada\u00adtion pilot should include enough episodes to sup\u00adport a deci\u00adsion, not to train a pro\u00adduc\u00adtion pol\u00adi\u00adcy. Most teams size it so they can mea\u00adsure accep\u00adtance rate per task and per oper\u00ada\u00adtor, and run one quick train\u00ading san\u00adi\u00adty check. That usu\u00adal\u00adly means a small, fixed batch per task, with a stop rule if the accep\u00adtance rate stays low after the first ses\u00adsions.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Three ques\u00adtions set the num\u00adber. How many tasks are in scope? How many oper\u00ada\u00adtors do you need to com\u00adpare? And does your ML team want to train a quick base\u00adline on pilot data? Pub\u00adlished sys\u00adtems show that small, clean demon\u00adstra\u00adtion sets can be infor\u00adma\u00adtive: the ALOHA authors trained fine biman\u00adu\u00adal skills from real demon\u00adstra\u00adtions col\u00adlect\u00aded with a low-cost tele\u00adop\u00ader\u00ada\u00adtion inter\u00adface.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">The PILOT readiness check<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Score each item from 1 to 5 before the first ses\u00adsion. Any\u00adthing below 3 is a rea\u00adson to fix the plan, not to start cap\u00adtur\u00ading.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1600\" height=\"940\" src=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__03-pilot-check.webp\" alt=\"Radar chart of the PILOT readiness check: platform locked, inputs logged, limits defined, operators trained, threshold set.\" class=\"wp-image-285\" srcset=\"https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__03-pilot-check.webp 1600w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__03-pilot-check-300x176.webp 300w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__03-pilot-check-1024x602.webp 1024w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__03-pilot-check-768x451.webp 768w, https:\/\/www.graveiensai.com\/blog\/wp-content\/uploads\/2026\/10\/collect-teleoperation-data__03-pilot-check-1536x902.webp 1536w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\"><figcaption class=\"wp-element-caption\">Score your pilot before the first ses\u00adsion.<\/figcaption><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><b>P: Plat\u00adform locked.<\/b> The robot mod\u00adel, end effec\u00adtor and con\u00adtrol inter\u00adface will not change dur\u00ading the pilot.<\/li>\n\n\n\n<li><b>I: Inputs logged.<\/b> The log\u00adging spec (streams, rates, clock, meta\u00adda\u00adta) is writ\u00adten and test\u00aded on a dry run.<\/li>\n\n\n\n<li><b>L: Lim\u00adits defined.<\/b> Accep\u00adtance cri\u00adte\u00adria, safe\u00adty lim\u00adits and stop rules are agreed in writ\u00ading.<\/li>\n\n\n\n<li><b>O: Oper\u00ada\u00adtors trained.<\/b> At least two oper\u00ada\u00adtors have fin\u00adished prac\u00adtice ses\u00adsions on the cho\u00adsen rig.<\/li>\n\n\n\n<li><b>T: Thresh\u00adold set.<\/b> The go\/no-go thresh\u00adold (for exam\u00adple, a tar\u00adget accep\u00adtance rate and a pass\u00ading base\u00adline check) is writ\u00adten down.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><b>Want the work\u00adsheet ver\u00adsion?<\/b> <a href=\"https:\/\/www.graveiensai.com\/contact-us\">Ask us for the PILOT readi\u00adness work\u00adsheet<\/a> and we will send the tem\u00adplate our team uses on scop\u00ading calls.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What drives the effort and cost of a teleoperation pilot?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Effort in a tele\u00adop\u00ader\u00ada\u00adtion pilot is dri\u00adven by robot access, rig build or cal\u00adi\u00adbra\u00adtion, oper\u00ada\u00adtor train\u00ading time, task com\u00adplex\u00adi\u00adty, the log\u00adging spec and QA depth. The biggest swing fac\u00adtor is usu\u00adal\u00adly hard\u00adware: ses\u00adsions on your own robot are lim\u00adit\u00aded by its avail\u00adabil\u00adi\u00adty, while a matched rig can run in par\u00adal\u00adlel but needs cal\u00adi\u00adbra\u00adtion against your plat\u00adform.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><b>Robot access.<\/b> Shared lab robots lim\u00adit ses\u00adsion hours per week.<\/li>\n\n\n\n<li><b>Rig work.<\/b> Build\u00ading, cal\u00adi\u00adbrat\u00ading or adapt\u00ading a leader arm or VR inter\u00adface.<\/li>\n\n\n\n<li><b>Task com\u00adplex\u00adi\u00adty.<\/b> Long-hori\u00adzon or biman\u00adu\u00adal tasks need more prac\u00adtice and pro\u00adduce more rejects.<\/li>\n\n\n\n<li><b>Log\u00adging and sync.<\/b> More streams and tighter sync tol\u00ader\u00adances add set\u00adup and check\u00ading time.<\/li>\n\n\n\n<li><b>QA depth.<\/b> Frame-lev\u00adel review costs more than out\u00adcome-only review.<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">Mistakes that sink a teleoperation pilot<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li><b>No writ\u00adten hypoth\u00ade\u00adsis.<\/b> With\u00adout one, nobody can say whether the pilot passed.<\/li>\n\n\n\n<li><b>Chang\u00ading the plat\u00adform mid-pilot.<\/b> A new grip\u00adper or con\u00adtroller resets oper\u00ada\u00adtor learn\u00ading and inval\u00adi\u00addates com\u00adpar\u00adisons.<\/li>\n\n\n\n<li><b>Log\u00adging video only.<\/b> With\u00adout joint states and com\u00admands on the same clock, demon\u00adstra\u00adtions can\u00adnot be used for imi\u00adta\u00adtion learn\u00ading.<\/li>\n\n\n\n<li><b>Keep\u00ading prac\u00adtice runs in the dataset.<\/b> Ear\u00adly ses\u00adsions teach oper\u00ada\u00adtors; they rarely teach poli\u00adcies.<\/li>\n\n\n\n<li><b>Skip\u00adping rejec\u00adtion rea\u00adsons.<\/b> A reject count tells you lit\u00adtle; a reject rea\u00adson tells you what to fix.<\/li>\n\n\n\n<li><b>Treat\u00ading oper\u00ada\u00adtor con\u00adsent as an after\u00adthought.<\/b> Ses\u00adsions record peo\u00adple too: their move\u00adments and often their faces or voic\u00ades. Use the same explic\u00adit-con\u00adsent process as any human-data col\u00adlec\u00adtion.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Tele\u00adop\u00ader\u00ada\u00adtion is one input among sev\u00ader\u00adal. For how demon\u00adstra\u00adtions com\u00adbine with ego\u00adcen\u00adtric video and sim\u00adu\u00adla\u00adtion, see our guide to <a href=\"https:\/\/www.graveiensai.com\/blog\/humanoid-robot-training-data\/\">humanoid robot train\u00ading data<\/a> and the <a href=\"https:\/\/www.graveiensai.com\/blog\/physical-ai-robotics-training-data\/\">phys\u00adi\u00adcal AI train\u00ading data pil\u00adlar guide<\/a>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">How Graveiens AI runs teleoperation pilots<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Graveiens AI han\u00addles oper\u00ada\u00adtor recruit\u00adment, train\u00ading, super\u00advised ses\u00adsions and deliv\u00adery for <a href=\"https:\/\/www.graveiensai.com\/teleoperation-data-services\">tele\u00adop\u00ader\u00ada\u00adtion data ser\u00advices<\/a>, with leader-fol\u00adlow\u00ader arms, biman\u00adu\u00adal rigs or VR-based con\u00adtrol matched to the clien\u00adt\u2019s plat\u00adform and bud\u00adget. Every engage\u00adment fol\u00adlows our explic\u00adit-con\u00adsent work\u00adflow (par\u00adtic\u00adi\u00adpant onboard\u00ading, record\u00ading and cap\u00adture man\u00adage\u00adment, meta\u00adda\u00adta tag\u00adging, QA), and clients are invoiced only for approved deliv\u00ader\u00adables. Graveiens AI is ISO 9001 cer\u00adti\u00adfied.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tele\u00adop\u00ader\u00ada\u00adtion is a new\u00ader offer\u00ading for us, and we are onboard\u00ading a small num\u00adber of pilot part\u00adners rather than pre\u00adsent\u00ading it as a high-vol\u00adume ser\u00advice. Teams that also need human task diver\u00adsi\u00adty can pair a pilot with <a href=\"https:\/\/www.graveiensai.com\/egocentric-video-data-collection\">ego\u00adcen\u00adtric video data col\u00adlec\u00adtion<\/a>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Frequently asked questions<\/h2>\n\n\n<div id=\"rank-math-faq\" class=\"rank-math-block\">\n<div class=\"rank-math-list \">\n<div id=\"faq-question-1761700000000\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">How long does a teleoperation data pilot take?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>It depends on robot access, rig readi\u00adness and the num\u00adber of tasks. Most of the cal\u00aden\u00addar time goes into the brief, rig set\u00adup and oper\u00ada\u00adtor prac\u00adtice rather than the record\u00aded ses\u00adsions. Plan the pilot around a fixed deci\u00adsion date and a stop rule, so it ends with a go\/no-go call instead of drift\u00ading into open-end\u00aded col\u00adlec\u00adtion.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-1761700000001\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">What data should be logged in each teleoperation episode?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>At min\u00adi\u00admum: syn\u00adchro\u00adnized cam\u00adera streams, joint states and com\u00admands, grip\u00adper or hand state, time\u00adstamps from one clock, and episode meta\u00adda\u00adta such as task, oper\u00ada\u00adtor, envi\u00adron\u00adment, out\u00adcome and fail\u00adure rea\u00adson. Agree rates, units and for\u00admats before cap\u00adture so the episodes load direct\u00adly into your train\u00ading pipeline.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-1761700000002\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">Who should define the acceptance criteria?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>The robot\u00adics or ML team that will train on the data should define them, with the data part\u00adner con\u00adfirm\u00ading they are mea\u00adsur\u00adable in QA. Cri\u00adte\u00adria should cov\u00ader task suc\u00adcess, allowed recov\u00ader\u00adies, dis\u00adqual\u00adi\u00adfy\u00ading events and sync tol\u00ader\u00adances, and they should be signed off before the first record\u00aded ses\u00adsion.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-1761700000003\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">Should we use our own robot or a matched rig for the pilot?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>Your own robot gives exact embod\u00adi\u00adment but ties up hard\u00adware time. A matched rig can run in par\u00adal\u00adlel but needs cal\u00adi\u00adbra\u00adtion and car\u00adries some trans\u00adfer risk. Many teams run the pilot on their own plat\u00adform to val\u00adi\u00addate the spec, then decide whether matched rigs make sense for scale.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-1761700000004\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">What should a teleoperation pilot brief include?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>A hypoth\u00ade\u00adsis, the locked plat\u00adform details, num\u00adbered tasks, accep\u00adtance cri\u00adte\u00adria, the log\u00adging spec, the pilot size with a stop rule, and a deci\u00adsion date. One page is usu\u00adal\u00adly enough; the point is that every field is filled in before the first ses\u00adsion.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-1761700000005\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">How do we know a teleoperation pilot succeeded?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>Com\u00adpare results with the hypoth\u00ade\u00adsis and thresh\u00adold you wrote at the start: accep\u00adtance rate per task and oper\u00ada\u00adtor, log\u00adging com\u00adplete\u00adness, and whether a quick base\u00adline trains on the pilot data. Clear rejec\u00adtion rea\u00adsons and a sta\u00adble accep\u00adtance rate are good signs that the pipeline is ready to scale.<\/p>\n\n<\/div>\n<\/div>\n<div id=\"faq-question-1761700000006\" class=\"rank-math-list-item\">\n<h3 class=\"rank-math-question \">Is consent needed from teleoperators?<\/h3>\n<div class=\"rank-math-answer \">\n\n<p>Yes. Tele\u00adop\u00ader\u00ada\u00adtion ses\u00adsions record peo\u00adple as well as robots, often includ\u00ading video of the oper\u00ada\u00adtor and motion data. Col\u00adlect explic\u00adit, doc\u00adu\u00adment\u00aded con\u00adsent cov\u00ader\u00ading AI-train\u00ading use, reten\u00adtion and with\u00addraw\u00adal, and keep con\u00adsent records linked to the episodes, as you would for any human-data col\u00adlec\u00adtion.<\/p>\n\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n\n\n<h2 class=\"wp-block-heading\">Scope your teleoperation pilot<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Know\u00ading how to col\u00adlect tele\u00adop\u00ader\u00ada\u00adtion data starts with a brief, a log\u00adging spec and accep\u00adtance cri\u00adte\u00adria. Send us your plat\u00adform and task list and we will help you shape a pilot billed only on approved deliv\u00ader\u00adables. <a href=\"https:\/\/www.graveiensai.com\/teleoperation-data-services\">See tele\u00adop\u00ader\u00ada\u00adtion data ser\u00advices<\/a> or <a href=\"https:\/\/www.graveiensai.com\/contact-us\">book a 20-minute scop\u00ading call<\/a>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Sources<\/h2>\n\n\n\n<ol class=\"wp-block-list\">\n<li>Zhao, T. Z., Kumar, V., Levine, S., Finn, C. \u201cLearn\u00ading Fine-Grained Biman\u00adu\u00adal Manip\u00adu\u00adla\u00adtion with Low-Cost Hard\u00adware.\u201d arXiv:2304.13705. <a href=\"https:\/\/arxiv.org\/abs\/2304.13705\" target=\"_blank\" rel=\"noopener\">arxiv.org\/abs\/2304.13705<\/a><\/li>\n\n\n\n<li>Alda\u00adco, J. et al. \u201cALOHA 2: An Enhanced Low-Cost Hard\u00adware for Biman\u00adu\u00adal Tele\u00adop\u00ader\u00ada\u00adtion.\u201d arXiv:2405.02292. <a href=\"https:\/\/arxiv.org\/abs\/2405.02292\" target=\"_blank\" rel=\"noopener\">arxiv.org\/abs\/2405.02292<\/a><\/li>\n\n\n\n<li>Cheng, X. et al. \u201cOpen-Tele\u00adVi\u00adsion: Tele\u00adop\u00ader\u00ada\u00adtion with Immer\u00adsive Active Visu\u00adal Feed\u00adback.\u201d arXiv:2407.01512. <a href=\"https:\/\/arxiv.org\/abs\/2407.01512\" target=\"_blank\" rel=\"noopener\">arxiv.org\/abs\/2407.01512<\/a><\/li>\n<\/ol>\n\n\n\n<h2 class=\"wp-block-heading\">About the author<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Graveiens AI Edi\u00adto\u00adr\u00adi\u00adal Team. The team writes from day-to-day work run\u00adning con\u00adsent-backed data pro\u00adgrams for AI and robot\u00adics teams, includ\u00ading tele\u00adop\u00ader\u00adat\u00aded demon\u00adstra\u00adtions, first-per\u00adson video cap\u00adture and anno\u00adta\u00adtion.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>How to col\u00adlect tele\u00adop\u00ader\u00ada\u00adtion data with\u00adout wast\u00ading robot time: a 5\u2011phase pilot play\u00adbook, accep\u00adtance cri\u00adte\u00adria and a go\/no-go score\u00adcard. Scope a pilot.<\/p>\n","protected":false},"author":1,"featured_media":282,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"wp_typography_post_enhancements_disabled":false,"footnotes":""},"categories":[4],"tags":[36,37,35,17],"class_list":["post-263","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-robotics-physical-ai","tag-imitation-learning","tag-pilot-scoping","tag-robot-demonstration-data","tag-teleoperation"],"_links":{"self":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/263","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/comments?post=263"}],"version-history":[{"count":3,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/263\/revisions"}],"predecessor-version":[{"id":286,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/263\/revisions\/286"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/media\/282"}],"wp:attachment":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/media?parent=263"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/categories?post=263"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/tags?post=263"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}