Download gemini_prompts.json from podolinsky/groot-libero-10-features-v3: direct link, hf CLI and curl.
- Browser
- Download file 10.5 kB
-
https://huggingface.co/datasets/podolinsky/groot-libero-10-features-v3/resolve/main/gemini_prompts.json
- Command line
-
hf download hf://datasets/podolinsky/groot-libero-10-features-v3/gemini_prompts.json
-
curl -L -o gemini_prompts.json https://huggingface.co/datasets/podolinsky/groot-libero-10-features-v3/resolve/main/gemini_prompts.json
10.5 kB
| { | |
| "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n", | |
| "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n", | |
| "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n", | |
| "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n", | |
| "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n" | |
| } | |